Tested the model briefly with my usual eval: a couple of questions on general knowledge most small models often get wrong, then write a fully-featured todo list app in JS, then rewrite the same app in Rust with Tauri. Granted, most models are well trained on basic todo apps, but it gives me an idea of the basic SWE capabilities I can build on. As far as I'm concerned, if it can successfully setup a local git repo, write a todo list app skeleton that works, I can work with it.
SWE: model is strong for its size. It one-shotted the Web app, had no bug. The Rust rewrite only had one bug (reordering didn't work immediately - fixed in one prompt). Committed locally then pushed to my GitHub (https://github.com/DexterLagan/RusTODO). Can't complain. If it can do that reliably, I can use it to make whatever I need.
General knowledge: I usually ask 2 questions many small models get wrong: summarize Operation Trojan Horse by John Keel, give publication year. Summarize the Ariel school incident of 1994. Qwen3.8 got the first question right, along correct publication year, gave glorious detail on Keel's theory, but got the second wrong. It thought that school was located in the USA. Ah well.
Performance on an admittedly overpowered laptop: 15 tokens/s in power save mode on this MacBook M5 Max 48GB, and 30 in performance mode. Perfectly usable for local coding through OpenCode.
Verdict: very nice local and free backup to my usual GPT/Claude/DeepSeek for code. Good for Web searches via Brave search tool calls. What more do you want from a small local model?
> General knowledge: I usually ask 2 questions many small models get wrong: summarize Operation Trojan Horse by John Keel, give publication year. Summarize the Ariel school incident of 1994. Qwen3.8 got the first question right, along correct publication year, gave glorious detail on Keel's theory, but got the second wrong. It thought that school was located in the USA. Ah well.
I'm not sure that this means anything. You're asking a ~27GB file to have losslessly compressed the entire training set (which apparently is a large chunk of the entire internet). That's not possible. Whether it happened to encode these particularly obscure facts losslessly or vaguely isn't really telling you anything about how good a model it is.
Like most models it makes up things when it doesn't know. I had high hopes for Gemma4, which was said to having 'solved' this particular problem - it didn't. Gemma4 made less things up, did say it didn't know more, but it's still far from perfect. By comparison Qwen3.8 knows more, but still makes things up when it doesn't know. Coding abilities are very impressive however, and yeah these SVG tests do seem to 'scale' or generalize over its general reasoning+coding abilities - at least in JS and Rust. My next test will be to ask it to write some macros in Racket, just to see if it can balance parens. Most, if not all models cannot, no matter their size.
Agreed. It's just to see how much knowledge the model still has for its size. For comparison, no other model of this size has ever given me the correct answer to the Keel question. Even GPT-OSS-120B got it wrong, but it's expected for its relative old age.
I concur, treating models as question and answer machines and judging them on recall is meaningless, unless you're measuring quantisation impact on a foundation model maybe.
100% agree, it's really just for kicks&giggles. The fact that the model answered correctly, unlike every other model of its size before it, still pleasantly surprised me.
Meaningless yet fun fact: DeepSeek V4 Pro 0813 made a much worse icon for the same app, and only produced an SVG I had to convert manually to .png. Qwen3.8 made a perfect icon in .png. I don't yet know how it did it, but it did it.
Qwen3.8 also seems to know French quite a bit better than Copilot, at least on common expressions. I have yet to run more tests for languages, but I'm baffled by its finer accuracy on the few sentences I asked it to translate, which were botched by Copilot. Go figure.
Credit where it's due. Qwen 3.8 27B is only the second local model after Gemma 4 that managed to correctly reason through one of my private benchmarks. It took 5x as many tokens to do it and 12m30s with MTP enabled, but it did do it.
Gemma 4 reasoned through it more implicitly, while Qwen 3.8 reasoned more explicitly. Laguna and Muse Glimmer failed hard on it, though they're useful for other tasks.
The VRAM usage seems way less efficient than Gemma 4 or Glimmer though, with 32K of context taking 2.5GB of VRAM. With those, even with MTP or a DFlash model loaded, you could still fit 256k-768k of context. With Qwen 3.8 27B I can't even fit 128k if I quantize V to Q4_0. Maybe with some trial and error I can find some settings that perform well enough with a larger context window that it's still useful for longer tasks.
Lots more testing to do, though I was getting some decent results out of Muse Glimmer which was more than twice as fast and supported huge context windows, managing to solve some bugs that Gemma 4 struggled with. I can't even begin to throw that task at Qwen, because just the prompt alone would use the entire context window and then it would reason for probably that same amount.
If you've got a 32GB card, it should be a decent model even if it really is memory hungry.
EDIT: Tried a few kv cache quantization settings, but it failed with those. I designed this benchmark to be pretty brutal in the face of KLD and any reasoning quality loss, so it's not too surprising. Gemma 4's QAT held up pretty well, at least and could consistently complete it.
Can you please tell which Gemma 4 variant managed to correctly reason through your private benchmarks? Was is Gemma 4 31B?
What quantizations and context lengths did you use for Gemma 4 and Qwen 3.8 27B?
I am asking because I can't even load Gemma 4 31B on my GPU with any reasonable quantization (even with small context), while I can run Qwen 3.8 27B with large context and good quantization...
If you don't explicitly point to the template file, then llama.cpp will either use the template inside the model file or it will use its own template copy and your results may vary. Obviously some of the template fixes are useful to people, so it depends if you're having problems with tool calling or can't fix the tool calling in other ways for your scenario.
My experience with the QAT models was that quantizing v to q4_0 gave me better results than q8_0 or even f16. I think the Gemma QAT models may have been QAT trained to expect a q4_0 quantized v cache. If you're not using a QAT model, I would leave both at f16.
Another thing aside from using the QAT models and a Q4_0 v cache since you're having trouble fitting the models, is that you don't have to use the mmproj if you don't intend to use vision. If you need vision, but are hurting on VRAM, then you should be using --no-mmproj-offload. That will keep the mmproj loaded in system RAM instead of on your GPU. Loading images will be a little bit slower, but it can still be quite fast and you'll have more breathing room on your GPU. If you don't provide the mmproj file on the command line at all, then it won't load it anyway. If you're using some program like LM Studio, a simple thing you can do is move the mmproj and mtp files out of the directory for the model so LM studio can't find them and then it won't load them at all.
For Qwen 3.8 27B, doing any quantizing definitely hurt results a lot, so in my case I used:
llama-server --model "Qwen3.8-27B-UD-Q4_K_XL.gguf" --spec-type draft-mtp --spec-draft-p-min 0.35 --spec-draft-n-max 2 -fa on -np 1 --jinja --ctx-size 65536 -b 768 -ub 768 --cache-type-k f16 --cache-type-v f16
So long as they have flash attention enabled, Llama.cpp enables Sliding Window Attention by default for Gemma 4 models. Even if they're using Ollama or LM Studio I would expect those to mostly be doing the right things.
Qwen’s 3.6/3.8 27b actually has some algorithmic advantage when it comes to the kv cache size needed, so it actually needs less memory at equivalent context. My experience using both in vllm supports, with considerably more overhead in context size on these models than Gemma 4 31b, and better performance in most tasks I’ve tried on both models.
Qwen 3.x does have an advantage but it's relatively small (64KB/token vs 80KB/token) - Gemma4 actually has less % of full attention layers, but the largest geometry and has the biggest "fixed" state for it's non-global layers. Muse Glimmer actually has by far the lowest per-token cache usage for the competitive 30B-class dense models - it's at about 13KB/token - very aggressive GQA (32Q/2KV) and also by far the smallest QKV dimensions.
Actually perf (speed) is going to mostly on token output, and here Qwen 3.x historically tends to lose badly as it tends to overthink a lot. I'll be running evals on 3.8 myself this weekend to see how its reasoning levels perform.
I was quite impressed by Muse Glimmer, and while I am sure people will observe that it is less good on benchmarks, my first experiences with this new 27B have been somewhat exasperating, whereas testing Muse Glimmer was rather fun. I have not tested either in an agentic context, mind you.
Yeah, Glimmer is excellent. You don't really test Glimmer with one-shots, because it's explicitly designed for multi-turn solution finding. The way I see it, if I've got a task that could be done either agentic or requires a lot of context (for example, dumping 600KB of API documentation and another 300KB of codebase for a project) then I would reach for Glimmer easy and it seems like it could get there most of the time.
Qwen might be useful to bring out for a second opinion on some more focused details that are largely information complete. Like, use Glimmer to bring together all the relevant critical data and evaluate what the actual problems are, then maybe prototype a solution. If it's still acting up, maybe throw the resulting context at Qwen and let it meditate on it.
I think there was some study done where ideally you would want to throw a bunch of different models at a problem since they don't all have the same perspective or diagnosis on what the problems or the solutions are.
That is exactly how this model has worked for me so far. Muse on a one-shot task will get to 80%. And if you even nudge it and say, "Hey, finish up," or "Review the syntax," boom, it's done. And I'm getting 20 t/s with Ollama on a MacBook M5 Pro with 48GB of RAM. It is a seriously impressive little model.
Glimmer works really well as an "explore" agent model (like in Opencode.) It seems to be extremely efficient at searching and collating that info, and executing commands.
From my testing so far, Qwen 3.8 is better at code but it tends to meander and take forever if it has to look in a lot of places. Glimmer will use like ~1k tokens to formulate a plan and Qwen 3.8 will routinely go over 10k
I wouldn't find glimmer interesting except that it has much less memory usage per token of KV than Qwen. So I can get 24x concurrent glimmer on 2xRTXA6000 (with 128k context) where I can only get 6 Qwen 27b. This means I can get something like 4x the aggregate tokens/s out of glimmer.
For some usages that speedup more than makes up for it being inferior to Qwen intelligence wise.
> The VRAM usage seems way less efficient than Gemma 4 or Glimmer though
Maybe it's implicit that you're using llama.cpp (although you don't mention GGUF), but it's hard to reach concrete conclusions about the model architecture based on one implementation in one runtime.
There's a tiny bit of play, like sliding window attention. As tokens leave the sliding window you can keep them or discard them. If you keep them, you can freely truncate the context and resume generation from an earlier point. If you discard them, you have to recompute the KV cache up to that point.
Llama.cpp checkpoints and moves snapshots of the cache to main RAM for faster resumption after truncation.
Have you tried Muse 30B yet? I have been impressed with it. I have Qwen 3.8 27B hammering away right now against Muse. And Muse is doing a little bit better.
#1: it does not require deep world knowledge, because that's not what local models are for.
#2: it directly attacks drive-by understanding, overly linear processing training, poor attention mechanisms, poor reasoning patterns or lazy assumptions that ignore very easy low hanging fruit.
#3: it requires solid instruction following in the face of errors. a lot of models will run into errors and then fall back into some kind of error recovery process that bypasses instruction following.
#4: does not require prompt fine tuning to tweak to each individual model. they all seem to understand.
#5: not unfair. almost every model demonstrates in their reasoning that they have the necessary information that if reasoned about appropriately, could arrive at the correct answer.
#6: not designed to add unnecessary complication that it is intended to exhaust reasoning budgets of any sort, so it is not inherently unfair to models that reason a little more or less. for example, it does not require unnecessary reasoning soaks (ie: hiding the prompt inside base-64 encoding)
#7: has real world use and is probably applicable to overall ability to generalize.
#8: can be scaled up as models get better.
#9: is a very good indicator of how bad a model is falling apart under various inference settings.
I especially like #8. If you have some free time (don't we all have so much of that?) it would be really interesting to run a binary search on each model you have, to see at what size/complexity level it manages to solve the problem, say, 50% of the time.
Bicycle is the right shape. Pelican beak is excellent. Nice background. Most importantly, the pelican has one leg on each side of the bicycle - that's very rare.
(No chain on this bicycle though - in the reasoning trace it says "already chainstay... skip chain detail; maybe a small chainring.")
It took 21 minutes(!) and used 22,276 reasoning tokens to produce 3,223 tokens of output.
(For the "they're training on your benchmark now" crowd, all of that cheating didn't prevent it from spending 20 minutes thinking about the task first! You can see the reasoning trace in the link I shared.)
For anyone who followed yesterday's Gemini 3.7 Flash pelican which rendered in Safari but not in Firefox or Chrome... https://news.ycombinator.com/item?id=49289112#49290012. - that turned out to be my fault, not the model.
Perhaps this isn’t a new observation but the problem with LLMs is very clear with these. It’s a nice microcosm. The LLM will draw a fish companion (unprompted!) with a nice gradient but won’t get the pelican’s feet right.
It’s obviously a problem of fundamental understanding and demonstrates that reasoning is more “directionless rigour”.
I'd like to see a human create a better pelican SVG without being able to look at the results. LLMs are language models, not (natively) vision models*. That's why it's a good benchmark: it engages LLM's logical reasoning in a way that we can check visually. The fact that they make errors which can be spotted visually, doesn't prove that LLMs are so far behind humans on language/logic tasks.
Besides which, "fundamental understanding" isn't a binary. Humans can also have a deep understanding of a subject and still make errors. (I do, anyways.) I do not think "fundamental understanding" is itself well enough understood to say definitively that LLMs do or do not have it.
*Some models have multimodal capabilities but these are separate weights, I believe it's only engaged in processing images. From the reasoning trace, the model never rendered the result -- only instructed the user to do so. I welcome correction if I'm wrong, I only have a surface-level understanding here.
Or a matter of having limited time but spending it thinking about the wrong thing. Which, not to anthropomorphize it, but it's not like humans don't do that, or at least I do.
It is interesting that it went with something like:
"pelican = ocean bird = nautical theme = this means I should put a captain's hat on the cartoon Pelican"
I think in some sense it understood the absurdity of a pelican riding a bicycle as intentionally nonsensical, so it went with a cartoon theme for the design and the shading. Worked out really well.
What's the result if you ask it for a "pelican equipment case" ? I've been trying the anti-bicycle pelican on some LLMs and the results are much more varied than the bicycle prompt. Some have very different ideas of pelicans (you'll get a small case with one DSLR camera in it, or a long rifle case, etc). You'll also get cases that are isometric view, or flat plane view from the front, or open or closed.
That is pretty good! It's interesting it chose orange, because that's actually a somewhat rare color for pelican cases. I would say 95% of the ones going around the world right now are either black or FDE (flat dark earth) tan colored. There are some orange ones, for sure, if you google image search "orange pelican case", people use them for emergency/first aid supplies or electronics. But only the smallest ones are available in orange.
It is also worth noting that it shaded the pelican case like it was a product photo taken in a studio, there's a virtual light source above and to the top left corner of it, then shading deeper color of orange as the body area progresses towards the lower right corner.
Doubtful? Without knowing anything about how it was trained, if you google image search "pelican bird" they are predominantly white in color. Though the bill is certainly orange, different shades depending on camera and lighting conditions. The SVG attempt looks more like an attempt at the product photos if you search "pelican case orange".
Or to make it even more tricky, since Pelican cases will float when closed, something like "generate an SVG file of a Pelican bird floating in the ocean next to a Pelican case".
I've been getting ads in my Instagram for http://colorcase.com. They add colors and wheels to cases and I'm sharing them because you're right, most of them aren't orange. But they could be!
Wow, those are expensive. I've been buying pelican 1510 and similar from eBay and b&h used/refurb. The harbor freight pelican knock off that is their 3800 and 4800 product numbers are also not bad. It's like 85% as good as a real pelican for 40% of the price.
Not too bad, on my 5090 I got ~160tokens/sec and it took 3m20s to generate a pelican using Ninfer which someone else mentioned. I don't know why I didn't get any sweet sweet seagulls in my background, even on xhigh reasoning effort.
Have you thought about running a second tier of the Pelican benchmark where you see which model makes best pelican on lowest or no reasoning settings? I think that'd be pretty interesting and might help highlight which models have a baseline capability - even within those I'd imagine the token usage would vary wildly and might give some indications on verbosity as well.
I've done a few variations, I've been impressed with all of them. My favourite so far has been "Generate an SVG of a turtle flying a kite", result: https://imgur.com/a/bdKJPV4.
Some will say conflating flying and flying a kite is bad, but I like it.
I was just updating a Proxmox cluster from v8 to v9 and had time on my hands and 15-30 mins after the models appeared for general download I had an initial pelican effort in .svg format.
I switch off reasoning by default and for starters I ran 3.8 the same way as 3.6 on vllm on a DGX Spark box. Open web UI prompted.
My word! It looks rather decent. Much better than previous efforts.
It's definitely trained on generating SVGs. Not that they trained on the pelican prompt itself, but probably a more general RL step that has it generate SVG and a stronger vision model judges it.
this sort of test surely only works when it's a brand new concept/prompt. like the exact same models trained newer data will surely get better at putting this together over time
There's a real change (compared to 3.6) in the way it writes in thinking — it drops words like "to" and "we" in "We need to", talks generally in note form, drops the/and all over the place, avoids "for".
"Need be helpful concise", "Need maybe not overdo", "Need ask!" Almost caveman.
I have an (unsourced, vague) suspicion that this rather unique thinking trace pattern is actually hobbling the MTP predictions, which seem to perform poorly.
Other notes: it uses the trick of repeating the prompt in the thinking trace.
It also worries about hidden chain of thought appearing in the final answer. It talks about "desired oververbosity 9", which is new. A bit GPT-ish.
It is being extraordinarily thorough in thinking through one of my code requests, but I don't know if the net result will be any better than the 35B MoE.
I asked it to ask me clarifying questions — it did, and it offered me a list of defaults I could simply agree to.
I don't think it is necessarily overthinking in the looping sense, but it is in the being exhaustive sense. I need to explore how it does with a tighter reasoning budget.
I am impressed but I am definitely in Camp Please-35B-A3B-When? here, because on an M1 Max this isn't really practical. I hope they do one, though I think they may not.
This is pretty much Chinese grammar with English words. In Chinese you'd say "need ask" or "I need ask" if you want to be unusually verbose, but there's no "to".
All your examples align with this too. It just sounds to me like its optimising by removing the superfluous English words and thinking more Chinesely.
I also noticed that if you ask it to write text content to a certain word length (like "please write 250 words about the potato"), it now takes its sweet time "manually" counting each word in its draft output. It puts a number next to each word and iterates through all of its draft paragraphs and tries to stick to a strict word count. This is with temperature 1.0 as recommended.
Annnnd the code of my WP code test is not better. It is bushy, overcomplicated, and has gone around the houses to do stuff it would not need to do if it hadn't overthought.
Oh dear. I need to try to understand what is going on here.
Compared to closed weight (especially unreleased and access-limited) and open weight/source large sparse MoE LLMs/VLMs, open weight/source small dense models benefits public the most because they just reaches more people.
Compared to Qwen 3.6, 3.8's thinking style changed drastically. With xhigh budget, it thinks a lot MORE, and longer thinking session directly translates to better performance. This tradeoff between performance and computation, memory, etc. is meaningful to me.
However, because Qwen 3.6 and 3.8 share the same architecture, with 32GB vram, llama.cpp, IQ4_XS model, MTP and FP16 mmproj, I can only get 200k context, which is not good compared to 640k context of muse glimmer. Hopefully this problem will be solved in Qwen 4.0 release.
If anyone else is running this on an RTX 5090, https://github.com/Neroued/ninfer as inference engine gets me ~138 tokens/second, roughly double what I get with a naive llama.cpp setup.
I'd love to understand this more. Are you saying the Qwen team spends their very impressive human and compute resources on publishing these amazing models and then botches the chat template with mundane bugs?
Like maybe I just misunderstand what's the hard part but wouldn't you assume that people who can put together an impressive model can also write a proper jinja chat template for it?
One really would think these companies (including Google) who spend many millions of dollars on compute could write a few hundred lines of Jinja correctly, so their investment works optimally or at all.
But they don't.
Then a couple of individuals on HuggingFace fix it, either a 2-person startup like Unsloth or a volunteer like froggeric.
I also don't understand how this repeatedly happens.
I did SFT / RL post-training on Qwen3 models a bit. This is an issue that dates back long ago.
My favorite theory is that they had many model variants internally, each using a slightly different chat template, so when it comes to the release they are not even sure what to use any more.
Yes for the first question. Google of all companies didn't even get it right with Gemma for a while until recently. For some reason it doesn't seem like people can actually get these templates right.
It’s unlikely they are benchmarking that downstream. They probably have private benchmark scripts with purpose-specific run telemetry and logging, etc.
I’m sure there is some basic testing but it might be agentic (LLM likes its own output) and maybe just some human smoke tests.
they likely use their internal infra to run benchmarks; aligning external releases with internal environments is always painful and somewhat underincentivized
The chat templates are usually the first thing that every major release bork on, and all new model architectures end up having a ~2 week initial window of small fixes before they’re not DoA
Temperature, top-up, top-k, min-p all control which token the model predicts next and how likely it is to select one token over the other.
You might understand this as "The capital of France is..." and the model isn't always going to select "Paris". Sometimes it will start a descriptive sentence or even get the answer wrong.
That selection of the next token is what these settings control, and lots of sub-optimal selections compound over time to produce a junk response.
Top-K: example setting 20. Select only from the 20 most likely tokens.
Top-P: example setting 0.9. Select tokens whose probably accumulates to this number. So say you have tokens with 0.7 then 0.2 then 0.1, the last will not be selected because the first two tokens already accumulated to >=0.9.
Min-P: example setting 0.05. Don't select tokens less probable than this value. So a token with 0.1 would be considered, a token with 0.01 would not.
The purpose of all of these is to exclude very unlikely next tokens.
I've run the inference to get the answers I linked to. If someone else does the same thing, that involves extra energy. If I read your conversation instead of generating my own, then that's one less tree that has to be chopped down.
Until we go advanced geothermal or we crack fusion, energy is dirty. Read my inference or link me to yours so we don't boil the planet.
Diverging from the sampler used in RL training is not good for long multi-turn results-- it's a great way to knock models into reasoning loops that wouldn't otherwise.
Depends on your tooling and quant? I grabbed the unsloth Q3 and it works out if the box in opencode. I had issues with OpenWebUI with a random 3.6 A3B.
If the benchmarks don't lie, this is getting very close to Opus 4.6 capability - which was the turning point for me for when AI was "good enough" that it became very hard to justify not using it.
I'm sure there's some benchmaxxing going on, and some things you get only with a a larger model.
But I'm feeling pretty confident if not by Gemma 5 than by mid 2028 we'll have local models that are almost always as good as Opus 4.6 was and in many cases far better.
Asking it factual information[1]. You just can't compress the entire human knowledge into a 30GB file.
[1] Without searching the internet. And even if you allow it, you'll get much worse results because search means browsing and parsing the top results, and search results are horrible, whereas internal knowledge from training encompasses the entire internet plus all books including very niche stuff.
> You just can't compress the entire human knowledge into a 30GB file
Fortunately, that isn’t necessary! What LLMs need is a level of fluency with key concepts so that they can (1) make effective use of retrieval tools and (2) understand the material in the context window. 30GB-sized models can absolutely store enough knowledge to do this.
Here is an example from the field of law. Most lawyers who have litigated contract disputes in California know about Civil Code § 1717, which makes any contract providing for attorney fees to a prevailing party mutual, so that even if the contract was written to be one-sided, it won’t be enforced that way. It’s a simple enough concept, but there are many more particulars to it, such as what happens when the fee provision is only written to apply to part of the contract. (Answer: it depends on other facts.)
When a lawyer recognizes that they’re in a situation where § 1717 is relevant, the first thing they will do is pull the statute and read it, because nobody has it memorized. And they don’t need to.
A downside is that you can't just download a lot of that knowledge, vs with the weights the copyright infringement has been outsourced to the lab. Nor can you just search for the info because the internet as a whole is increasingly aggressive at blocking anything that looks like an AI agent.
I'd love to see more retrieval powered local AI-- I think it's an area that open source development could excel. ... but there are advantages of having the knowledge in the weights!
Perhaps what needs happen is for someone to make an "ultrapedia", an AI restatement of a huge library of reference works-- created expressly for the purpose of being a locally stored corpus for AI agents.
Amusingly enough, someone from a frontier lab could probably answer this empirically. Years ago Microsoft demonstrated training LLMs on synthetic text-- books rewritten by LLMs to be more concise and more accurate. https://arxiv.org/abs/2306.11644 It is well known that Anthropic extensively uses synthetic text in training. You could probably get good 50tile, 90tile, 99 etc numbers just from the size of the training materials on Anthropic servers.
On a more serious note, it depends on your cutoff for "entire human knowledge". It's easy to prove for a generous interpretations of "entire human knowledge" that it can't be done, but hard for something like "all useful human knowledge".
IDK, 30GB is a lot of data when we're talking about text!
Moby Dick, uncompressed, is ~1MB. Compressed, it's around 500KB.
I feel fairly certain that one could fit all of the textual knowledge required to cultivate a world-class <insert name of preferred professional knowledge worker> in <60,000 Moby Dicks. (Arguably in <5,000 Moby Dicks with intense effort/pruning).
I think a specialized model could squeeze all you need to know about a certain profession in 30GB. But not all professions at once, which is what these models try to do.
Or maybe not, maybe there's a world model needed for human level at any profession that is very hard to quantify and requires more than 30GB by itself.
Yea to be clear I think >70% of the information is not profession specific.
I just think about all the content I’ve consumed in my life to become a professional software developer and I would be very surprised if it couldn’t be adequately represented by <30GB of uncompressed text. Most of the work was in “training”, not really in data.
The “foundational overlap” of K-12 education is identical for most professions with 2-8 years of “finishing” on top.
My mental model is that the budget is spread across maybe 20% genetics (most of our instinctive/genetic information is surely pretty useless), 50% k-12 education, 30% for professionally-specific knowledge.
The human genome is 800MB, (and 500MB with compression according to GPT) so that part is easy.
I think the problem is that the conversation drifted from "all useful human knowledge" to "enough knowledge to do my job".
Both a human and a current large model will bring up unprompted some tax regulation that applies in your district that could matter to your task. A small model won't know the tax code of every city in the world, as it's probably impossible to fit, and it's the kind of thing that unless you already know about, it's very hard to search for, let alone come up with unless specifically prompted about something tax related.
Unless we start having different small models trained for a certain profession for a certain city, I think we'll need those trillion parameters.
IME using 5.6 Luna and DS V4 Flash, I notice that although they are excellent at programming, even Opus-like in the way they try to debug, the thing they are worst at is inferring user intent and making good decisions with little information. They are absolutely terrible at that, will misinterpret small wording ambiguities. I suspect that's an ability you can't add with RL training, that it requires the depth of understanding from vast pre-training.
Similar to the way they asked Sol to solve Erdos problems, that's what I want my model to do for programming.
I don't want to try to take my best educated guess at what the best design is BEFORE implementation - especially if you're designing a feature for a codebase you're not an expert in, you don't know like the back of your hand (i.e. one that is mostly or entirely LLM generated).
What sounds good on paper - often times becomes unideal in practice when you get to the reality of implementation.
It may not be worth re-architecting your entire system to get to a "pure" design that would be the best - all things considered.
Instead, I'd like the model to independently design many plausible and coherent good solutions, then implement each of them, then intelligently pick the few winners (after its fixed any bugs that could be causing promising solutions to look artificially bad) - unless there's an obvious one - and then give me the data I need to make an informed decision on which one to go with, all before I even look at the design or implementation.
You're not getting this from a one shot prompt from a 30B model today. You can't even really get it from Sol or Fable - IME. But you can get somewhat close.
Even Fable is bad at this, I would constantly have to fix it going down architectural dead ends or just making obvious mistakes.
Which sucks for people that want LLMs to do everything like a genie, but does mean senior engineers have a few more years before they become redundant.
I think someone ought to encode this into a harness. It is really insightful into how we should be spending time if it is going to be spent reviewing AI code.
This is a harness problem not a model problem, try prime agent it can do that and it will do it well even :P but you need to prompt it in according to its tools and processes.
Overall I'm very impressed with how well this did. It's a big improvement over 3.6, and it feels on-par with some much, much larger models. I think this one is on-par with Gemini 3.7 Flash.
One thing to note - the build for this on my RTX 6000 pro blackwell took a long time. Easily one of the longest builds I've done. It took around 2 hours to build the site. Obviously we'll have some quants for this soon that will accelerate things, but I was still surprised with how long it took.
Getting 30 t/s on a Mac M5 Max laptop. You should be getting close to 200 if you can use the 5090 acceleration tools I see posted here. There are forks for the 4090 and the 3090. Maybe there's a fork for the 6000?
You should be getting way more than that on a 6000 pro even today. I'm getting 40tok/s on a pair of 3060s. You can ask a SOTA model to optimize your setup for you.
One thing a lot of people don't seem to factor when hyping Qwen is how much models like this tend to 'overthink' with seemingly endless 'second guessing'. 3.8 seems no different from what I've tried thus far.
As capable as it is, it's hard to justify using it when a competing model (e.g. Gemma4:26b-a3b) can consistently achieve the same or similar response with only 1/10th as many 'thinking' tokens, achieve much higher tokens/second, and take a small fraction of the time. I suppose 'YMMV' depending on your use case.
Also, I haven't used it enough yet to see if it's prone to infinite looping, but its predecessors sure were.
IME this is a strong/reliable model smell, you typically see smarter and less benchmaxxed models' thinking traces spending more time exploring the solution space, and benchmaxxed models more time trying to refine/decide on the response contents. It has always been a big problem with qwen.
In general it's kind of a benchmaxxing/distillation (I don't mean that pejoratively here, read on) artifact where test time compute's purpose is to refine/zero-in on a particular input:output pair. Basically they're trying to "remember"/re-derive the answer rather than arrive at it deductively - it gets comical/absurd when you see it expending a huge amount of tokens trying to figure out the answer to some trivial question or response to some input, as if it were a trick question or the choice of wording was of the utmost importance. But it's not a bad thing when the output matters or the task is hard: basically the model has been trained to respond/act like a much smarter model and it's probably better for it to overgeneralize that behavior.
Also, IME Gemma and most other local models that support thinking tend to have the same problem, and it's partly only a "problem" because they give you access to the thinking tokens themselves (which you don't in many cases see when working with frontier american models) and you can actually see how they're getting spent. For the local or bulk (runnin on owned or rented hardware) use case where you are paying in time/watts rather than purely by the token, IMO it's a good idea to just look at the task success rate / time per task and not worry about the raw thinking token count.
Given that it apparently defaults to 'xhigh', this is probably the answer.
Granted, it's still much lower tokens/s than you'll get out of many MoE models.
Edit: Even set to medium or low there's still a lot of second guessing, less consistency, lower 'acceptable response' rate, and slower/more token churn vs gemma4:26b-a3b. I think gemma4 is just a better 'general purpose' model.
I haven't tried lowering thinking, however, I actually asked a solid question earlier regarding a real world scenario I encountered and all that excessive thinking made it give me an amazing answer. The thinking actually all made sense, and honestly I found it thought of similar stuff to what I thought when I drew my own conclusion.
I don't usually rely on AI for much (I'm actually kind of anti-AI, although I follow stuff like this enthusiastically because of the rapid advancements, "average joe" access, and openness), however for what I asked? It was spot on.
The subject was a bit personal, so I won't share. I was just curious what AI would say about the situation and it definitely surprised me, especially since Qwen, while usually great on development/coding stuff, has shown weaknesses in other areas.
Definitely a solid release, and this one runs on my 4090 with minimal loss of quality!
In my experience with Qwen3.6 35B-A3B, disabling thinking made the model generate inaccurate replies. Ask it for the recipe of egg salad and it gives you the recipe of an omelette.
Did I miss something, is it possible to have that model be reliable without thinking?
Turning off thinking mode you need a different tokenizer and a whole host of settings. It's more like switching your vehicle from gas to diesel, than turning your headlights from lo to hi.
“Thinking” is just a guiding methodology to help iterations (between the initial prompt, results, and a mixture of harness back and forth to the LLM) converge on something sane in a massive parameter space.
I like to think of it much like (as a common example most people can relate to) the Newton-Rhapson method for finding roots of a (mathematic) function. Your initial prompt runs, then the ‘harness’ kicks in using whatever methodologies are behind them to iterate on that prompt (back and forth with the model, occasionally with the user to get better guidance) and refine the outputs to hopefully converge back to some sensible output or actions the user was initially looking for.
So you’re hoping for an LLM that sort of ‘zero shots’ or needs minimal iterations from a prompt to give usable results. I find from my anecdata it varies across models and what I’m trying to get it to converge on. I tend to prefer models to not zero shot attempt because they tend to not do great, I want them to get feedback often to let me push them down the route of convergence in spaces I already understand well, meanwhile I like them to explore and give me new paths in spaces I’m not too familiar with.
That’s really what all that “second guessing” is, it’s making sure you’re following a sane path in a massive parameter space of an ambiguously defined problem. Imagine if in Newton’s method you checked the slope and it didn’t decrease from the last iteration and you just say “screw it let’s keep trying that direction.” LLMs and their harnesses tend not to have that base assumption like iteration on decreasing slopes to guide them closer to convergence, it’s a lot messier.
Use 3.6 27b as a daily driver for months with charmbracelet crush. Gemma 26b-A3b is not even remotely comparable in terms of coding for me. YMMV depending on how you work, what harness you use, etc I suppose.
Yeah. I use it to do extensive work on full source bootstrapping, deterministic operating systems, compiler debugging, kernel debugging... all with one tiny go binary without endless NPM deps like pi (which /I/ regard as trash)
Works better than opencode (pi based) or anything else I have tried for my needs, and by far the prettiest and easiest to reason about what is going on.
But I will bite. What does pi do today better than crush for your use cases?
Wow. This model is so good, and we have GLM 5.3 (seems great voor security related work) and Deepseek. In a few months we'll have Fable/Sol-like capabilities that are not coming from the big US companies. I feel as a programmer that that is more than enough. How wil OpenAI and Anthropic survive when frontier model intelligence becomes commoditized?
I stopped using Fable and Opus 5 because I literally can't understand the output. The waffle is so intense it no longer makes sense. I don't understand who is using those to get real work done.
I'm using kimi-k3 on a real high level dev & analysis task at the moment (with Gemini flash for fast implementation) and it is fantastic.
I think we have passed the point where frontier intelligence is commoditized.
just curious; which fable/opus versions are you using?
Fable-1m-max-thinking output, despite being incredibly slow and expensive, feels like it bucked a trend towards superficial loquaciousness in their models that had been building since 4.6.
I'm a bit of a luddite when it comes to upgrading models, fable was the first one to make me give up Opus 4.6-1M-max.
People will claim it's not comparable to Opus despite it beating the score. I'm not sure I disagree, but I'm also unsure whether I care. Most new models nowadays are "good enough". I cannot complain because I'd rather spend that time improving my prompts and docs. Opus might be a _slight bit better_ at picking up vague hints, but it's also extremely expensive, and I hit the 5 hour limit way too quick.
I care a lot about speed and efficiency right now. For my setup I would like to have 2-3 different model families. I've settled on GLM-5.3 (formerly Deepseek v4 pro 0813) for architecting, Deepseek V4 Pro 0813 for developing, and Gemini flash lite (any recent cheap model) for repo scouting. I'll add another one in the mix for reviewing (in this case Gemini 3.7) and that's all I need.
I've tried most models except Grok.
Qwen is too expensive IMO (Alibaba Cloud subscriptions are hard to come by and I'm not spending 50 euros a month for a tool, so 18 euros it is). If it ever becomes efficient enough to run locally I will definitely look back.
Claude is slow and expensive (the cache hit prices are absurd).
OAI is pretty good, I might add it to my arsenal seeing how cheap it is.
These opinions change every day. Last week I would've never picked Deepseek until I read about the pricing. even post aug 16 it's worth it (although it's getting close to gemini pricing).
Right now my costs are 12 euros a month (z.ai) + whatever deepseek consumes. This typically isn't more than 8 euros a week. 44 euros a month and I have a setup that is doing pretty well.
Grok 4.6 is a game-changer. I have yet to go back to other models after starting to use it. You just can't beat the price + output quality (even K3 is more expensive)
I'm convinced a lot of the anti-open-weight model comments at this point are inorganic traffic - there's trillions in investor money riding on a world where these models aren't cheap commodities. Having actually used things like the recent GLM, Kimi, and Qwen I think any edge the labs have is marginal at most and actually prefer the open weight models in most day to day usage.
Anthropic's recent releases are wordy to the point of exhaustion. Every time I use opus recently I find myself wanting to yell "GET TO THE POINT" at a terminal, which is exacerbated by it being slow.
Why are you convinced of that? Pretty much every time I see statements like that online, I can find plenty of organic traffic supporting it, not everything is a bot.
I just wouldn't bias myself that way, most people haven't really used local models. This stuff is pretty much all subjective evaluation, there's plenty of reasons for people to favor certain models or disfavor others.
I'm generally extremely skeptical about a lot of the model hype that show up in comments. Except when there is an extreme mismatch the performance, quirks and quality of these things are difficult to nail down. You wouldn't know that from the comment section of every single release.
I think some of these are excited, eager users always ready to hype up the new thing. The same crowd that previously would constantly push for a rewrite from angular->react->svelt->god knows what. Instead now it is on a 6 week cycle and about models/harnesses.
I think most of it is bot driven spam by the various labs. It's hard not to notice 3 month old accounts with very strong opinions about various frontier labs and little else.
Ultimately, I think some of it is legitimate shifts in who's in lead and what is the best. You gotta dig through a lot of crap to get to that, and I don't really know how to to.
Ultimately I'm saying is that I always applied a fair amount of skepticism about what I see in comment sections but these day it is extreme amounts.
> I think most of it is bot driven spam by the various labs. It's hard not to notice 3 month old accounts with very strong opinions about various frontier labs and little else.
I've assumed the same as well.
I also assume that many of the companies developing these models engage in benchmaxxing.
At my company we've developed our own internal benchmarks for evaluating LLM models as they become available. The benchmarks are tailored to our particular use cases but the utility and knowledge our benchmarks assess is still fairly universally applicable. I see wide differences between what our internal benchmarks report and what the major benchmarks do.
There was a whole lot of fanfare about how amazing GLM 5.2 was when it was released, but it was pure rubbish on our internal benchmark -- far behind OpenAI, Anthropic, Gemini, DeepSeek, etc. I don't know how to reconcile the fact that GLM 5.2 performed very well on some of the major public benchmarks, but consistently performs so poorly on ours. OpenAI models tend to dominate our internal benchmarks.
It actually is showing in public benchmark if you know how to look for it. For example, in Terminal-Bench 2.1, GLM 5.2 received 78%, while GPT 5.6 Sol received 88%.
Then Terminal-Bench 3.0 came out (where the questions are new), and GPT 5.6 Sol received 34.6%, while GLM 5.2 dropped to a whopping 4.6%.
> OpenAI models tend to dominate our internal benchmarks.
That's odd, since Fable seems to be the leader for the industry. Not cost-effective, but if Anthropic models get dominated by OpenAI in your internal benchmarks, this calls their validity into question. Separately, see the jagged frontier effect. [1]
I've been using GLM 5.2 at my day job (mostly Rust backend work ATM). Nothing that blows away the models from OpenAI and Anthropic, but solidly good enough to get it done. A lot of people have experienced this and the fact that an open weights model can do so is where most of the excitement comes from. Optimizing for benchmarks can only get you so far, and people are quick to criticize models that fall into it (like DeepSeek Pro V4 recently).
I think most of it is bot driven spam by the various labs. It's hard not to notice 3 month old accounts with very strong opinions about various frontier labs and little else.
Really no different than when there was suddenly online personas everywhere hyping up TSLA out of the blue. You can see the same thing going on with the BoringCompany subreddit. Crazy that the botnet master isn't able to convince us that Grok is also the best model. I don't think buying twitter was an accident it was probably just literally covering up the evidence.
IMO this is getting hyped because the 27b version runs on a decent gaming GPU. This is NOT a thread for their largest model, this model will run on a mid-high end gaming PC, which you probably have in your household. Mine is 6 years old and it runs quite well.
I strongly suspect that many of these accounts you think might be bots from the labs are just people who only have one interest.
One thing I have noted a lot more of is that there is comparative fanboying going on. Like "this model has done badly, my favoured competition has a model out soon that will beat this in every way" — comparing a released product to unverifiable hopey claims about an unreleased product.
It's tempting to assume that is bot stuff, but if you've been around any other "hot" technical hobby online (cameras, phones, 3d printers, whatever) you will know it's not. It's just fans aligning into teams, some of them laconic and amusing, some of them overkeen and toxic.
Which itself released yesterday? You're writing, reading, and evaluating enough software in a ~36 hour period to form, reject, and form another opinion about which model makes better _architectural_ choices?
Honest question, how do you assess models this quickly? What metrics are you using? Would love to get my suite from multiple days and hundreds of prompts down to minutes. Got a few first pass tasks I run upon release for an initial experience, but those only work because even Fable and Sol fail despite objectively correct solutions existing, so it works because most models fail, but then, those are consciously not enough for coding, tool use, adherence or task specific inference and assessment…
Anything from ML pipelines for language specific pruning over a Rust/JS/CSS mix codebase to assistance in motorcycle maintenance and different canvas coatings. Most of my evals build on those requirements and especially past failures, whether in pure information, task execution and coding or tool calling beyond the overfitted mainstream. All stuff derived from actual failures encountered, some still only few models come even close to passing. With such a mix, it just takes a while to get any serious opinion on a model. Doubt anyone can do that in such short time, unless their tasks are so simple that most modern models not only succeed but could themselves accurately rate output. If even Fable still confuses PU or wax coated cotton canvas with a nylon shell, or tells me with a straight face to adjust the valves on a bike that has hydraulic lifters that needs experience for human assessment and the time that comes with it. Anyone with less knowledge either wouldn’t see the mistakes staring them in the face and just go by vibes, any model rating these equally can not tell what is accurate and will just go by the output sounding accurate over being. Gives sometimes very interesting results far different to public benchmarks. Inkling, e.g. is more accurate in not telling you to adjust valves that are simply not adjustable then Fable or Sol, which just tell you to adjust every 5000km. Sometimes even when their reasoning and search includes sections about the fact this is not necessary or possible. The beauty of overfitting and unbalanced training data…
Most of us aren’t deploying these in general purpose use-cases. E.g. I use Qwen mostly for vision in my personal assistant. I have an eval set for that. Pretty much each of my use cases has a pre-computed problem set.
Likewise I have a tool-use eval set and a browser-use eval set. I use frontier models for most interactive tasks that are not home assistant or background agents.
I can imagine someone could build evals for that but I have never done so.
no they aren't. they discontinued their always-sold-out coding plan and launched QwenCloud (basically a friendly frontend with Alibaba Cloud as the hidden backend) and launched typical subscription plans for Qwen & co alongside it.
With max reasoning, Luna is actually less than half the actual cost to run compared to DeepSeek V4 Flash 0731 with updated prices (based on Artificial Analysis Cost per Task)
They were significantly more expensive for coding compared to DeepSeek's offering because of much higher cost of cache reads and lower cache hit rate. Now, after the price jump, the effective cost of deepseek.com API vs 3-rd party providers is about the same.
I think it's very clear that someone who has checked out all the models but the one that called itself mechahitler and is explicitly being fine tuned to support far-right politics is making the choice for reasons other than performance and cost. It's not like all the other models even had plausible claims to those metrics.
As I said to a dead reply, for coding all of that is immaterial, as long as it codes well then that's all that matters to most people, except it seems those who have an idelogical issue in which case the other model companies also have issues.
No, coding does not make ethical issues immaterial. Nor does it justify your behavior of harassing someone to do something that they've clearly decided not to do for ethical reasons on the basis of criteria that they have clearly rejected.
I don't care how good Grok is, it will be a cold day in hell when I give a guy who goes to AfD rallies, instrumentally supported the current Ebola outbreak, and gives the Hitler salute on stage (don't try to tell me he did not, I saw the video) any money I'm not absolutely forced to.
And yes, OpenAI and Anthropic are highly questionable, too, and so I've ended my OpenAI subscription and I'm moving from Claude to other providers. But even people who do not do that are absolutely justified in drawing a line at supporting Musk, because there's a difference between mere normal shittiess and Musk.
The funny thing is it doesn’t even code well. It’s still worse than OpenAI, Anthropic, and the leading open source models.
So you get worse outcomes while also financially supporting one of the most prolific pushers of race wars, the destruction of labor rights, environmental protections, and general misinformation.
Truly a win-win?
Contrary to your framing it seems to only reason to use Grok at this point would be ideological.
There are a lot of people who are apathetic to what musk is, people that don't care are not people who should inspire you. What the hell is so inspiring about apathy anyway?!
And yeah, people that don't care DO make the world worse through their apathy.
I would say it is sad that there are people who use Grok when there are so many other choices available which don't come with the issues of supporting Musk.
It is not all just 'politics'. Take a stand on some issues. It doesn't cost much not to use Grok.
> No one actually cares about the politics as long as the model codes well.
I do. There are enough models to choose from that I don't need to use one from a guy who did a nazi salute at the presidential podium, and proceeded to rid our government of brown people he felt didn't deserve their job.
i care about not financially supporting a person that is actively trying to disenfranchise me, why is that a difficult concept for some people? that not everyone is motivated exclusively by financial profit? is moral bankruptcy so pervasive that some people assume it is unanimous?
'They' - in this case - are those who are unable to empathize. People who only learn of other's struggles through visceral personal experience. Conservatives, in general, struggle with this class reasoning... Very much a 'if I don't experience it then it doesn't exist' kind of mentality. Frustrating and exhausting. A lack of imagination, a lack of empathy, they just are lacking in these regards.
Your comments are all in absolutes. As if this was the one thing that mattered to you, everything else is irrelevant. Don't most people make their decisions based on a balance of considerations? Dismissing half the world as nazi or fascist does not seem productive, it kinda makes you look like the insane one.
I hope really badly that we'll get a new 35B A3B or similar MoE model!
I also miss the Qwen 3 Coder Next, which was 80B A3B, there are quite a few use cases where a non-dense model <100B would be the sweet spot (when you have the VRAM but not the TDP or compute power). Heck, I'd gladly take A5B or A8B or even A10B as a sort of middle ground.
I'm hoping too that they'll put out some MoE variants.
Qwen3.5:122b:a10b can run about twice as fast as this 27b dense model.
Edit: Like its predecessors, 3.8 seems really inclined to overthinking, and on a 27b dense model that's kind of painful.
I think I'm going to stick with gemma4:26b-a3b as my go-to because it runs about 4x as fast and tends to only need a fraction of the tokens in its 'thinking' stage to get the same or similar answer.
Probably helps it score a little bit better in benchmarks :) `medium` seems like a nice balance so far; along with some light steering to vary think effort as needed for task and being pragmatic.
Interestingly 'medium' is the closest thing the _model itself_ has to a default thinking level. The chat template injects directions [1] at the very start of the system message when the reasoning effort is 'xhigh' or 'low' but 'medium' implicitly just means no added reasoning-level instructions.
[1] The specific directions are "Reasoning effort is set to xhigh. Please think carefully through the task, validate key assumptions, consider plausible alternatives, and prioritize correctness, consistency, and clarity in the final answer." and "Reasoning effort is set to low. Keep your thinking brief and focused, moving directly to the conclusion without unnecessary elaboration."
"inclined to overthinking," holy cats you're not kidding! On a Mac mini M4 Pro 64GB I prompted it with "svg owl" and it thought for 17m12s, outputting 36.3KiB of thinking chatter. It did end up producing a 20.2KiB HTML+JS+SVG file with a very nice owl, including cursor-tracking animation, but it ran for more than a half hour!
I think that's called test-time scaling i.e using more tokens at infer time to squeeze out higher model performance. That's must be part of the explanation for good benchmark results.
that seems to be how most Chinese models achieve increased benchmark scores. GLM and Kimi models are "thinkslop" models that reason over their own thinking, which increases cost and decreases speed significantly. That's why GPT is in a different tier altogether - faster, smarter, and sometimes cheaper.
Yep.. it's pretty obnoxious for real-world use with the default 'xhigh' thinking. Ridiculous amount of "Wait, actually.." which might help for complex coding tasks but makes it unbearable for general purpose use.
No agentic harness; I just prompted literally "svg owl" into LM Studio 0.4.21+2. macOS Tahoe 26.6.1, Mac mini M4 Pro (14 core, 20 graphics cores) 64GB RAM. (I put the machine config so the 17m12s of thinking has some context.) It got about 12.75 tok/s and used 21,769 tokens.
For comparison's sake, Qwen 3.6 35B A3B MLX, same prompt, same machine, thought for 1.59 seconds, used 2,398 tokens, and hit 80.83 tok/s. It actually output just a straight SVG file of an owl (not a bad owl either!), whereas Qwen 3.8 27B spent a zillion times longer gold-plating everything and output HTML+JS+SVG with interactive animation, sound, etc., etc.
Did you try the claude reasoning traces finetune for qwen3.6? I find that it works muuch better. I assume the same 3.8 finetune will be released at some pointas well.
AgentWorld is a pretty recent MoE release from the Qwen folks with quite a bit better performance than 3.6, released around the same time as 3.7 Plus/Max, I guess. Not sure why it didn't get more attention, as it is definitely better than 3.6 35B A3B on all dimensions, but especially for agentic use. Still nowhere near the dense models (even 3.6 27B), but clearly an upgrade in the small Qwen MoE line.
KAT Coder is another Qwen 3.6 MoE fine-tune that also improves on 3.6 by a measurable amount.
But, I do hope for a bigger MoE, in the 70B to 120B range, something in the Coder Next lineage. I've got a Strix Halo that isn't getting used to its best ability because the best models all run fine on my faster desktop dual 32GB GPU setup, and they run too slow for comfort on the Strix Halo.
And, yet it outperforms Qwen 3.6 35B A3B on Terminal Bench and SWE, etc. I dunno.
Edit: I guess you're right; apparently it's for simulation. I didn't look into it beyond the benchmarks. But, it does work in an agentic context, regardless. It'll write code, and drive an agent.
Same here! Qwen3.6-35B-A3B is the only local model I've found that runs reasonably on my iGPU. Looks like me and and my noisily-wheezing laptop will be sitting out this upgrade.
[self-reply because comment edit window ended]: I now have a head-to-head benchmark. On my wheezy laptop (specs in sibling comment), this new dense model, Qwen 3.8 27B, gets ~4 tokens/second on generation. The older mixture-of-experts model, Qwen 3.6-35B-A3B, gets ~20 tokens/s.
MoE is literally 5x times faster (on CPU) than comparable dense Qwen!
So you happen to know how to demonstrate this higher capacity, what do we mean by that? Can understand bigger codebases? And how do we explain it, what part of the model does it?
Don't mind! It's 64 GiB dual-channel DDR5-6400, i.e. roughly 100 GiB/s of bandwidth. (AMD 7840U (Zen 4))
I'm using a Q4 quantization from unsloth (Qwen3.6-35B-A3B-UD-Q4_K_XL). It gets up to ~20 tokens/second in generation. I don't know precisely how much KV cache I can safely use, but it's in between 140k–256k. (I.e., 140k reliably works, 256k kernel-crashes from OOM. Don't feel like bisecting).
Inference is llama.cpp with the Vulkan GPU backend on Linux. (I.e., -DGGML_VULKAN=1 on the llama.cpp build, and --gpu-layers all on llama-cli or llama-server. (And for my specific setup, two kernel parameters specific to amdgpu: ttm.pages_limit and ttm.page_pool_size. A driver VRAM limiter. Look it up if you're on amdgpu!)).
Thanks! I don't have any knowledge of running models locally.
I assume it would not be able to handle an unquantized Qwen3.6-35B or is it irrelevant as you almost always would want to run a quantized version of the model on consumer hardware?
not parent, but 4-bit quantization is generally consider a good trade off for speed/performance, so you might use it even when you aren't on consumer hardware, but definitely when you are on consumer hardware.
Me too. 35B A3B runs really fast on my MacBook Pro (M4 Max) and is suitable for real-time tasks like dictation post-processing. The dense model is not.
Total param count decides how much vram you need to run it. Active param count decides how fast it runs. My 10 year old GPU can load quantized 35B or 27B, but it can’t process 27B parameters per token faster than 2-4tok/s, while it can do A3B at >40tok/s
I'm still confused about Qwen 3.6 35B A3B. Everything I read said that the 27B model performs better at coding tasks, so what's the purpose of the 35B model?
Also radically better on an M1 Max. I get well up into the 60s t/s with the A3B, stuck at 9.5 or so with this new 27B, though perhaps an MLX build will help.
The "a3b" refers to its active parameters -- unlike 27b it is a mixture of experts model, so it runs much faster, about as fast as a 3b model, but needs as much memory as a 35b model! So good for unified memory systems like macs :)
anecdotes: 35B-A3B does want more memory, bigger model. But if you get it running it will be faster and more enjoyable to use -- text will fly by -- due to only 3B params being active, in my experience at least.
35Ba3b is usable in plain cpu inference on a fast server, the dense model is MUCH slower. On GPU the 35ba3b is still around 2x the tok/s single threaded, which can be a good tradeoff for some applications.
MoE models have less active parameters in play at any particular moment, so they perform much faster on lower bandwidth memories (like Strix Halo or DGX Spark) and also use less memory generally.
Dedicated GPU memory tends to be much faster (either DDR6 or HBM). So if you can fit a whole dense model in it, you're probably better off with that.
I hope the bonsai team makes another 1bit quant of this model (or releases code/instructions on how to do it), using the Qwen3.6 27B on my 16GB mac mini has been wild . The 1bit quant feels like opus level… for the first couple turns. Then it has trouble eg switching from plan mode to act mode. This is mostly mitigated by starting a new session. (tbf this limitation is called out on the hf page)
I saw unsloth has 1bit quants too so I might check that out, anybody have experience with those?
nope just my normal m2 mac mini. I bought it just as a normal computer to do my taxes and whatever, so it’s mind blowing that I can run this kind of AI workload on it. Well, “run” might be generous, it gets like 3-5tok/s
I’m working on a setup that’s more geared towards running tasks overnight so the slow tok/s doesn’t matter as much
100k is all the context I have ram for, this is with any auto-compact turned off. This is using Cline in vs code. I’m sure I could tune the system prompt and mode switching more to work better with this specific model, but I haven’t gone down the custom harness rabbit hole yet.
And this is also specifically for the 1bit quant version. I don’t think the fp8 or even fp4 versions have this issue, but I haven’t tried those much
I wonder how this practically compares with Muse Glimmer, especially quantized.
I've got an RX 7900 XT (20GB of VRAM) and I can run glimmer with a full 128k context window with the draft model at 65-80 tok/s.
This model, on the other hand, I get about 30 tok/s with a 30k context. Raising the context or loading the draft layers for MTP drops performance to 9-15 tok/s.
So I wonder how big the "real world" delta between Glimmer and Qwen is here. I can already run 3-bit DSv4-flash at 9-15 tok/s with 100k~ context, and I suspect it would outperform 4-bit Qwen 3.8 27B here.
I'll have to experiment and see if I just made a stupid mistake somewhere, but it looks like Glimmer might make more sense for the comically specific niche of "20GB VRAM".
I'm in the exact same boat with a 7900 XT and a good Glimmer 30B experience. I was really hoping qwen 3.8 would bring some memory/space efficiency savings along the lines of whatever is going on with Glimmer 30B. I have been surprised that a 30 billion model fits and runs better (at higher unsloth quantization! UD-Q4_K_XL fits!) than a 27 billion model.
I'm not seeing that with that exact quantization from Unsloth, so far. I'm seeing a _lot_ of thinking before it starts doing, but it all seems pretty reasonable and not loopy (at least no more loopy than big models, with the expected "But, wait! I need to..." types of back-tracking). So, it's taking a long time, but I don't think it's doing anything pathological.
I would characterise it as obsessive, not loopy. It's definitely burning through a lot of tokens to ruminate about aspects of tasks that earlier models get done better seemingly on memory.
I still need to understand that.
Setting a reasoning limit does not seem to have good results, because it really seems to go down rabbit holes and that means that cutting reasoning off too early is going to punish the quality on anything it has not got round to pondering yet. But maybe I have to give it a bit more room.
I have not tested in an agentic sense yet, just with my sort of pet queries in LM Studio, but it rather looks like it expects an agentic flow, because telling it that it's a helpful coding agent and changing the order of things in my prompts (telling it up front to ask any clarifying questions before detailing the rest of the prompt) has definitely kept its thinking a bit more on track.
Since it might be helpful to some, here's my current commandline for llama.cpp running on an RTX 4090 with my monitor moved to the iGPU to free up all of its VRAM.
Identical to the qwen3.6 config. With a prompt like "svg owl" (which can reuse quite a lot compared with creative writing or similar, so ngram-mod shines), I get about 70-80t/s like this, with a memory overclock of about 1.5GHz
In my tests, even Q8 quantization for the KV cache comes with notable drops in performance for longer tasks. It does provide more context length in limited RAM budgets, but the longer context tasks are where KV quantization starts to show problems. It’s basically unnoticeable for simple and short tasks.
> --spec-draft-n-max 5
5 is a lot of tokens to draft. Are you really seeing acceptance rates to support that? When I tested it, 2-3 was the peak. Anything more started reducing performance except on highly predictable short outputs.
The thing is that I can either use the q8 context, or have not enough context window, so I just live with whatever degradation there is. The same can be said about the IQ4_NL. I would not go any lower though.
As for the draft count, indeed that depends on what you do with it, but for coding, reverse engineering and that kind of stuff it does pay off in my testing, though 5 is really pushing it, but the 4090 has so much compute.
Last logline I saw scroll by right now had 47% acceptance rate for 4th and 28% for 5th, but not sure how representative that is.
I think when tuning 3.6, I saw more like 33%? But not 100% sure.
I wonder if we could take a page out of the Solar power book and do MPPT but for draft count.
Constantly sweep through possible values to find the best result for the current conditions.
Though it might turn out that that doesn't offer any benefits. Has it been tried yet?
___
EDIT:
LLM pointed me towards this thing I'm not going to read
Don't you just need to occasionally try n+1 for your current setting? Then record acceptance rates. The nature of the thing is that if you try n, you already have a good sampling of how well m works for all m<n, right?
This is all a pretty active area of research, both adaptive drafting, different draft approaches (DFlash etc), I expect to see really big improvements land in the popular runtimes this year.
Okay, so, I did some actual coding with qwen3.8 in the last 2 hours, and if I (let the agent) take the average of all those "mean len" loglines, and let them get weighted by how much happened leading up to that logline, it's an average of 3.33
The unweighted average is 4.1
So with qwen3.8 as well, I think I would not go below 3 in this setup with this workload. Meaning 4, and, to push it a bit further, 5.
Also, cool that it carries over from qwen3.6 like that.
I have found that q8 kv quantisation is really bad in anything requiring some kind of literal retrieval even at relatively shorter contexts (5-10k). Eg it would mess up names, titles of papers etc that were deeper in the context.
How many tokens/second you getting? I have the same CPU but a 5070 and only 64 GB of ram. I just got llama.cpp built and am now hitting a whopping 5-6 tokens/s.
I absolutely love this comment. I wished there was a website where people would post their working command lines as well as what hardware they are using to run that stuff on + tokens / sec prefill + gen.
Problem with that is I think that it quickly devolves into cargo culting, nonsense and noise.
Arguably, what I am doing is also very very close to that, with the only difference being that I am somewhat less of an idiot than the average internet dweller you'd get on such a site. Or rather a different flavor of idiot.
Ideally, the people building the tools build them in a way that just does the right thing - which I am confident that llama.cpp does or will do in the future.
So you encode that knowledge not in language and online comments but in code and with a filter for actual expertise.
And, frankly, there's really not all that much to it. It's like maybe 3 parameters to play around with.
The valid solution space is pretty small, but people will want to make it "theirs" regardless, so you get non-solutions just so that everyone could also be a part of it. The usual social dynamics foo.
I've found that once you factor in multiple GPUs things can get complex quite quickly because the default packing routine in the LLM runners tends to be very coarse resulting in substantial amounts of VRAM wasted. More so if you start running drafters and multiple models at the same time.
Let me return the favor:
Deepseek V4 Flash 0731 Q8
This is on a 4x3090 box with 256G RAM and a 5955WX TR; the CUDA drivers are patched to allow the 3090's to talk to each other. MB is a WRX80E-SAGE (which has 7 slots at x16), BAR > 4G on and PCIe BAR resize set to AUTO in the bios. OS is Ubuntu 25.04.
I'm still working on an optimal packing for the draft variation on this theme but that gives all kinds of issues. The 1,000,000 context is very nice to have.
still could! there's the network site https://ai.stackexchange.com/ just need people to ask questions and to downrank accepted answers older than a year (that might not be possible though).
Thanks for posting! Have you had any success with running without kv cache quantization? Is there a noticeable difference in quality without any? I would assume that would eat into context but 170k is pretty generous!
According to this shitty vibecoded thing "I" built https://hypfer.github.io/will-it-fit-llama-cpp/ (and I guess according to math too), FP16 K/V would give me something like 90k context at the same model quant, which doesn't really fit my usage.
But maybe someone else has experience to share there
just to clarify. yes YOU built it. just because you used some tool doesn't mean the idea, prompting, reprompting, babysitting was not your creative input and effort.
put differently, if you put a random person infront of whatever model you used (say, a 50yo receptionist at a pharmacy in india), they would not have been able to create that, because they would have lacked the motivation, idea, background knowledge, taste, etc to create such a thing.
I've worked with a good few types of managers of the years, none would take ownership of my work. The end result? Sure, for shareholders/this managers that is fine.
But that's not the same as me being needed to be on a call with integration teams and having to run the call. It's just a layer of abstraction for management. I own the creation side of the product.
this topic always goes into this level of ridiculousness. Your question can't even be answered. We should focus more on encouraging discussion, not asking unanswerable questions. It wont bring anything other than points of both sides with no objective correct answer.
I generally agree and expect this to be the case from a legal perspective.
Legal questions of authorship are going to have to be established in terms of doctrines like SSO [0] and AFC [1]. Currently the incredibly sparse caselaw around this has yet to involve such non-literal notions of copyright.
So if I hire an artist and am a motivated individual, have an idea for a painting, have background knowledge about paintings and have taste in paintings and can offer a critique of the painting as the artist paints it, then somehow I created the painting?
Nothing absurd about that. What do you think an "Executive producer" is? A "Director" ? Does Peter Jackson get credit for creating the Lord of the Rings Trilogy films? Christopher Nolan for his films? But did he make them ? No, it was the collective effort of thousands of individuals all working under their direction.
Just like if somebody creates software today, and the end result is generated by the collective effort of thousands of agents, the "Director" still gets credit.
That’s an interesting way to look at it, and still generous to the llm.
By which I mean it puts the llm in the role of actor and such… people practicing an art. It seems harder to make a case that the llm is actually making creative decisions rather than aping a synthesis of past human decisions and expressions of taste, strictly, at every turn.
Though I guess it’s all “shades of grey”. I never felt like I was contributing a ton when I was essentially wrapping a large, capable library that some smarter person or people made, either.
Tangentially… remember back when we used to see headlines like “12 year old made a web browser!” and we all thought, “No. No they didn’t.”
What does it matter who receives more credit? The producer produces the film and the director directs it. If I pick my phone and record a video, I'm now the producer and director of the film and the sole creator of it.
If there were multiple people involved in the creation of a film I helped to create, I cannot factually say I created it. Just like if someone builds something using code generated by AI, they can't factually say they created it.
I think in that case it's fair to say you created the painting with the artist, even if the artist should get majority credit. I don't like the analogy though, to me it feels more like you're a project manager directing a team of genius but single minded interns.
I've never encountered a project manager who has said: I built this thing, after a project has wrapped up. I have encountered ones who've said I managed this project.
I've heard some variant of "we built this thing", which I think is valid. I'm not sure how I feel about anthropomorphizing a coding agent at this point though, which is what you're implying.
I just read through a couple of your posts that weren't dead or buried, and it seems like you're pretty anti-AI. You should really start to have an open mind towards it. It's going to be the future (if it isn't already), and as you continue to get older, you're going to really wish you spent your time right now learning and embracing the technology instead of being so against it. A lot of the skills and things that you're holding on to right now might not be relevant by then, but you'll be at a disadvantage from not keeping up with the industry and need to play catch-up.
> I'm not sure that typing messages to a chat bot requires much catching up
Lol, yeah I'd wager AI tooling skills are less relevant than knowing your way around a shell. It doesn't help that all of the AI tooling has event more churn than js libraries and package management systems!
I think there's probably some value in understanding how LLMs work, but beyond that it's the same boring skills that matter the most... Critical thinking, design chops, attention to detail, perseverance, etc.
He still didn't create the paintings, he offered instructions / feedback to the artists that created them. Whether or not he is credited with authoring them is an entirely separate issue.
There have been plenty of workshops where artists hire assistant painters while maintaining authorship over the works themselves, from Rembrandt to Warhol to Hirst.
Sure - and a factual statement would be, these artists hired assistant artists to help them create their paintings. Just like it would be factual to say that the person directed a LLM to build an app for them.
The person above is running the OS from a separate GPU, so the command can use all of the VRAM of the 4090.
If you're trying to use the 4090 for your desktop at the same time, you will need smaller context size. You could remove the '-c 170000' and try turning fit on to see what fits.
>But nvidia-smi only showed Xorg using 200M out of the 24G, so why would a 911M alloc fail?
That was just the last buffer allocation request that failed, it didn't tell you by how much it failed by. It could have failed it by a few kilobytes, it could have failed it by 910MB. One would guess it probably failed it by a couple hundred megabytes in the end judging by your results.
Does anyone know how to get this working with Claude Code via llama-server? I'm getting a jinja template error about the system prompt not being the first message.
You could run the 2-bit quantization at about 10GB for the weights. You'd probably also need to quantize context to get a decent-sized context. The Unsloth model card shows the 1-bit quantization doing seemingly shockingly well on a task (but there is notable loss in model quality for each bit you give up below about 6 bits). An MoE model can partially reside in system RAM without catastrophic performance loss, but a dense model like this one cannot. You have to run it entirely out of VRAM, or it'll slow to a crawl.
Yes. It won't be as fast as fitting the whole model into VRAM but llama.cpp defaults are pretty smart about GPU/CPU splits these days. Just YOLO it with `llama serve -hf unsloth/Qwen3.8-27B-GGUF:UD-Q4_K_XL` and it'll definitely at least run.
You may be able to run a quantized version. One strategy I've found effective is to set Claude or Codex loose on the problem and they'll do a decent job setting up the best possible model revision.
I can get 128k context on a 5070ti with 16 GB of VRAM (using the Unsloth 2-bit quant[0]). This is via a .bat file on Windows 11. I'm getting about 50-60 tokens/second and the quality is much higher than Qwen 3.6 27B. I'm using llama.cpp[1]:
1. download two files llama-b10434-bin-win-cuda-13.3-x64.zip and cudart-llama-bin-win-cuda-13.3-x64.zip from that llama.cpp Github releases page, and extract both into the same folder.
2. Download the Qwen3.8-27B-UD-Q2_K_XL.gguf file from huggingface and put it into the same folder beside the `llama-server.exe`.
3. Create a file named "RUN_QWEN_3.8.bat" next to `llama-server.exe` and put the text above into that bat file. Double-click the bat file, then open http://localhost:8033 in your browser to see a chat window.
You can use it with any agents by pointing them at http://localhost:8033/v1 which is a working OpenAI compatible endpoint (it doesn't use a token, if you give one it's ignored).
Congratulations, you're now running Qwen 3.8 27B.
Note: I built the computer in question for playing games, yes it needed to be Windows 11 for anticheat reasons to play games with family, I didn't want to dual boot so here I am. I figure I should share instructions for folks who may also have a Windows PC around for such purposes. Specs for this are AMD 9800X3D, 32GB of system RAM, RTX 5070Ti 16GB
For the things it does, what tool is better than ffmpeg? Really struggling to see the cargo cult angle. Similar for llama.cpp, as it is literally the only framework I can get to run on my multi-Radeon rig. It is the most portable runtime out there.
It depends on what model you’re running, and for what workload. For personal use (one or two convos at a time) with models that fit in gpu memory, pcie bandwidth doesn't really matter. Just try and be on gen 3 x8 or higher.
Llama is decent at auto optimizing it if you let it use both gpus. It’ll split the workload so the contiguous layers are all on one gpu. Once the model is loaded, you only transfer weights between gpus once (per token?), at the layer boundary.
I run on an 8gb 3070 and 12 gb 3060, and the only weird thing is that the weaker card gets more layers (and therefore work) because it has more ram.
Oh, if you’re barely fitting the models into your vram, you may need to explicitly adjust the layer balance between cards — sometimes it fails to realize it should have put certain things (like draft models) on the other card so you can fit one more layer in.
A hetero-GPU setup is definitely cost-effective if you don't strictly require the raw speed of a top-tier card like 5090. Just keep in mind that the total throughput will also be bottlenecked by the slower card.
To provide some anecdotal data, here is how my 5090 + 3060 setup performs with Qwen 3.8 27B (Unsloth's UD-Q4 with MTP):
As opposed to loading it up in RAM + VRAM? Pretty much always better to split it up to multiple GPUs. My priorities are load up all available VRAM, then offload MoE experts to RAM (if possible), then offload other layers.
I use an RTX 3090 (24GB) and a GTX 1080ti (11GB). Just the 3090 for 3.8 I get maybe 60tg/s (UD-Q4 quant), for both I get around 40tg/s (UD-Q6). Not apples to apples though considering it's different quants.
Yes. I can split a model like this across 3 GPUs (a 1080 with 8GB and two Titan Vs with 12GB), and it's much faster than running it on 36 CPU cores. As long as it fits in aggregate VRAM, it seems very advantageous to do so.
Depends on your pcie connection. If they're both x16 then it's pretty low overhead, x8 is ok, but x4 is too slow. Also it's a bit tricky getting an optimal setups with mismatched vram, I think you could probably still make use of the full vram if you're clever but it's trickier.
for layer parallelism (e.g. to get more vram) the bandwidth between layers is essentially nothing (like 16kb per token I think), so I don't think x4 would even be a problem!
I haven't tried this either but I'm guessing if you could pool the GPU memory over whatever the kids are using these days, I think it was SLI back in my day. The GPU memory should still be faster than the RAM?
That's a very deep rabbit hole involving PCIe topology on both the hardware and software (NCCL) side, among other things. It's too system-specific to answer directly, but the entrance to said hole can be found at https://github.com/local-inference-lab/rtx6kpro/blob/master/... .
Disregard references to RTX 6000 cards, most of it is generally applicable to all multiple-GPU boxes.
Lol at that command. Why is this stuff so hard to run locally? I've spent a few days trying to figure it all out and haven't been able to. LM Studio doesn't work behind proxies. Ollama is confusing and doesn't seem to support Qwen3? And Llama.cpp is your command.
I just want to run `<some-command> <model-name>` with some default parameters set and for it to run locally.
What makes you say that it would be hard to do that?
It's long, I guess, but not cryptic.
You tell llama server where the model is, which context size to use, what to use for the K/V cache quant, that it should do MTP, tune some MTP parameters, and that's kinda it.
Perfectly logical blocks with all the model-specific weirdness (that does exist!) abstracted away.
You could also just run -m <modelfile> and let llama-server do the right-ish thing. The defaults are probably fine, but not how you squeeze out these exact numbers.
I think at least. I've never tried. My hubris stopped me from trying auto configs.
> llama server where the model is, which context size to use, what to use for the K/V cache quant, that it should do MTP, tune some MTP parameters, and that's kinda it.
isn't that the hard part? You know the ballpark ideal values for these many parameters since you're a knowledgeable expert but the vast majority of people are just like "I want AI" and have no idea what all the jargon even means.
Why would they wish to handcraft this ? That is what agents are for ?
They could ask your current agent to a) search for this type of content online for the optimal setup for their hardware b) have the current agent/harness spin it up have it verify the config run few experiments.
Sure AI may make mistakes, or won't get the best possible config probably, but it certainly do a good enough setup, this is a task with feedback on whether the server crashed or poor performance easily measured so the agent can do a pretty good job.
Because this is kinda the one new thing that arrived in the technology scene, so getting at least some amount of understanding of its "inner" workings might prove useful in the future.
Beside that, it is also just.. interesting? It's fun tuning the machine to see it improve. For some, anyway.
The pain point they raised is this is too complicated for people who just want to get started, that is not true anymore.
It is certainly fun to fine-tune and setup if you like do something like that, however the need to do it hardly is a barrier for those who don't want complexity as OP imagines.
Lower level API/interfaces should not be a barrier for people if they are apply framing that way. More and more people are thinking agent native so this is not really a issue.
The alternative is to label it AI. That is distasteful for some of us, Intelligence is much stronger than automation in our minds so we avoid the term and try to describe it differently .
This is exactly it. I already have broad access to Claude, Gemini, GitHub Copilot. I want to use open models on automated tasks that chew up tokens but where I don't necessarily need the best-in class models and UX.
For Claude, I setting a single config file and then download and run Claude Code CLI. Even easier for the GitHub Copilot CLI.
I'll actually defend you on this one, but not only the command, that's more of a fine tuning option. On the hugging face link for example, there's no flat 'download' area. It's completely unintuitive for people to know how to use it. Reminds me of my first time trying to make sense of torrents when they first arrived.
There are some clients that will index the models and allow you to do that but I'm no expert, I've used OLama studio but it always seems to go weird for me.
Even this command above, it's not clear where op got the model from. So I'm with yah.
For example, op uses : Qwen3.8-27B-IQ4_NL.gguf.. But I cant see where to download it. It's not tagged on hugging face at least..
Official llama.cpp releases ship with huggingface support. If you don't want to download it yourself, you can just use the `repo/model:quant` convention and it will handle downloading locally for you.
But you're assuming I'm using the Olama studio. This model as far as I see doesn't have a gguf download.. Unless I'm missing something on the page.
If I want to download the model myself, it's not clear. I thought it was supposed to behave like a package manager. But even in nuGet I can download a zip of the package.
What on earth are you talking about? llama.cpp != Ollama. You can (and should) just use llama.cpp directly. Upstream llama.cpp can take the shorthand huggingface path and automagically download it into a cache folder as part of the launch. Have you read any of the docs?
Excuse me, but thats a direct link you've just sent. I asked where I can find the links. I like to believe in the source of truth.
They shared a lot of links, I'm struggling to find yours. Where did yours come from ?
Who is Unsloth AI? Have they modified the model ? Is this really the source of truth ?
Do you see how steep the barrier for entry is to do anything right ?
unslothai is not a name qwen has ever used. So you're sharing a link to a model that isn't from the owner, while saying it's the owner's. I'm not comfortable with that, and I want AI to be a better tool.
> Excuse me, but thats a direct link you've just sent. I asked where I can find the links. I like to believe in the source of truth.
You asked where to find the GGUF files of this model for direct download and I provided it. Almost all useful model files that can be downloaded are hosted on Huggingface.
> They shared a lot of links, I'm struggling to find yours. Where did yours come from ?
I went to Huggingface, went to the Unsloth org, as they tend to be the best, went to the Model page, and went to the "Files and versions" tab.
> Who is Unsloth AI? Have they modified the model ? Is this really the source of truth ?
Unsloth AI is a very popular, highly reputable organization that takes upstream model files, performs some optimization, and provides models in various formats. Apart from speed tweaks, they do not modify the models. They also provide useful benchmarks, copious documentation for local execution, and a Studio application for easy execution and post-training of models.
> Do you see how steep the barrier for entry is to do anything right ?
No. Searching for this information is not difficult. The llama.cpp documentation and guides that Unsloth provide are all you need. Search engines can take you further if you want.
> unslothai is not a name qwen has ever used. So you're sharing a link to a model that isn't from the owner, while saying it's the owner's. I'm not comfortable with that
Qwen also provides models in GGUF format on Huggingface, but they will not be as performant. Even when first-party GGUFs are available, most people will prefer quants from Unsloth or a few other popular optimizer accounts.
> I want AI to be a better tool.
Best of luck. Your attitude and unwillingness to even try and learn on your own when people have tried helping have burned my good will, and this is as far as I'm willing to carry you.
I gave you a direct link to where you can get a download, and explained why I chose to link you where I did. I can't post screenshots here drawing you a map of how to use a website. I don't work for Huggingface or any AI company and can affect no changes to how intuitive any of it is, and think is easy enough already.
Huggingface, Unsloth, and llama.cpp all have documentation you can follow that will exceed anything I can tell you here. LMstudio, Lemonade, or Ollama might be even easier for you to use. Take my suggestions or don't.
Start by copying the command line from the Unsloth guides.
You don’t need to fine tune all of those parameters to get started.
It’s really easy to ask an LLM to adjust the command line if you can’t be bothered to read the help out. Copy the help output into the LLM and tell it your goal.
> Ollama is confusing and doesn't seem to support Qwen3?
Typing “Ollama qwen3” into Google takes you right to this page:
If even Googling for basic Ollama support is too hard, there might come a point where you have to acknowledge that local LLMs are not for you. None of this is really that hard with some basic Google bootstrap skills or by asking an LLM to help with the command.
on Qwen's official docs: https://qwen.readthedocs.io/en/latest/run_locally/ollama.htm.... It's not like I just made it up. Of course I searched "ollama qwen3" and saw what you linked, but that doesn't mean it "works". I have other things to do besides to try a bunch of poorly documented and executed tools just to see if it works or not.
I guess the TLDR is that I'm stupid or lazy. Also, everyone is responding about how easy it is, and yet, it's apparently so easy that it's hard to document well.
There are easier ways to run it. OP seemed to enjoy tinkering and customizing the command to run it exactly the way they want. When I don’t want to tinker Unsloth Studio is probably closest to pick a model and voila.
Well there's a lot of knobs to turn if you want to improve performance. You can always point an LLM at the model card, give it your info, and have it write up the command.
Sure, but shouldn’t the programs to run the LLMs go “the user has this much vram and the model is this size, so I’ll start with sensible defaults based on that”?
> Every big company in the world uses a network proxy.
It's becoming more rare, now.
A lot of the universal truths about corporate networks from the early 2000s are no longer true today. Some companies are stuck in their ways though.
The overlap between companies that require someone to use a network proxy and companies that have GPU-equipped machines with enough RAM for LLMs and and that allow people to download and run executables of their choosing has to be small.
I would love to see more data on that because I've seen it constantly. There is more isolation maybe where you can do whatever on 'open' network, but always some kind of proxy/vpn connection for hitting anything sensitive.
The operlap is there.. But I would be worried if it was just flat out taken away from secure managed connections just because of AI.. Again, would love to see the numbers of your assumptions.
Woa TIL. I thought that was somehow long solved at the OS level or with VPNs or something like that (no idea exactly how, I'm sure just I'm misunderstanding something basic).
Makes it rather weird that LM Studio doesn't support it given how their target market, or well at least for their paid products, is very enterprisey.
I never install this stuff manually anymore. Just tell your LLM of choice to download model X from URL Y, build the latest inference engine of choice E, and then create batch files or shell scripts to run instruct and/or reasoning models in accordance with instructions at URL Z.
While you probably have the audacity to check the commands and code that your favorite LLM generates to complete this task, i'm pretty much certain that at some point in time, due to this pervasive mindset of "just let an AI agent do it," we'll have a front page news story of someone who had their family photos on their NAS uploaded to some public space, their security cameras connected to a "live porn for lolz" webring, and perhaps even their crypto wallets and bank accounts pilfered, all because their favorite LLM got exploited by some prompt injection attack while trying to be helpful doing system administration tasks.
I remember there was a short story in BYTE Magazine about a similar kind of scenario way back when, I think at least 30 years ago, long before LLMs and AI agents became a reality.
Hmm I'm still struggling with the downloading of all the models I want to try locally, so I might be out of the loop, but all of this, including tuning (to your local HW), feels like it could be automated. I'm old though.
WOW, my first try running on my 2 3090's, it was a bit slow... but it FEELS like opus 4.5, i gave it an image and a broad overview of what I wanted it to build, and it built the whole thing from beginning to end.
Beating or comparable to Opus 4.6 in benchmarks. Opus 4.6 was released in February, 2026. So if we still want to talk about a "6 month difference" between Chinese and American AI, the sentence should now be:
Chinese (small model) AI is 6 months behind American (largest model) AI.
I'm a huge open model fan, and have used them since forever, even have daily drivers for on-prem dev, but no. They do not beat opus on real-world usage.
Qwen models are impressively good for what they are, are "good enough" for plenty tasks, can be ran locally on decently priced hardware, and so on. They certainly have their uses, and the field in general has advanced faster than my early expectations. But to compare a 27B model to SotA behemoths from a few months ago is doing everyone a disservice, especially people who pick it up, try to use them just like API models, and leave disappointed and confused. Number goes up on a benchmark isn't it.
We have an internal eval that measures performance on tasks for a handful of embedded systems repos for our mmWave radios (mostly Rust, some C for microcontroller stuff). Qwen3.6-27B scores only 4% lower for pass@1, n=250 compared to Opus-4.8.
For the labeled dataset, the average PR size they're being measured against is around 1.5k SLOC.
This is very much "real-world usage" for us. The sort of change sets that come in daily/weekly and are solving non-trivial issues in the respective codebases.
As is usually the case, the most broad claims from both the labs and from the consequent pushback are talking past each other.
4% is within the margin of error anyways for pass@1, so I think pass@k > 1 is gonna be the better indicator of any movement (still need to calibrate the optimal k to re-test). 10 seems too tolerant even though that tends to be the next tranche I reach for.
Right I take your point, but that's not the comparison in this case though. These are paired results on the same 250 tasks, so whether a 4 point gap is meaningful depends on the task-level deltas, not just plugging p=0.04 into the binomial SE.
How much does it score though? 0% would be 4% less if Opus was at 4%. Unless you mean relative fraction not percentage points - but people usually mean percentage points in such situations.
> ...but no. They do not beat opus on real-world usage.
I agree, but then we just need meaningful benchmarks that clearly show that! Otherwise it's hand waving about something that should be put on paper in quantifiable terms.
If you are working in a company and using language models, it is a very good idea to hold a bunch of evals you can trust and use to validate new models. Calibrate every once in a while with prod data. We have our own and the only numbers on quality and cost I trust come from this setup.
That’s well and good, but how are we supposed to evaluate the accuracy of random HN comments without anything resembling somewhat objective metrics? People say all manner of things, and usually it’s contradictory. What heuristic do you propose?
> but then we just need meaningful benchmarks that clearly show that!
That's the rub. AI benchmarks are IMO, by and large totally unreliable. We think of them as similar to traditional benchmarks of deterministic processes where the number of variables is low. But they're anything but that. Non-deterministic processes with an astounding number of variables and fuzzy acceptance criteria.
It leads to results like these, where if you take it at face value, the only conclusion you can draw is "wow Anthropic must be stupid if Opus takes 1T parameters to do what Qwen can do in 27B."
The only useful benchmarks are those you've created for your specific workflow. Only then can you assess whether a given model is better or worse for what you are using it for.
Yep. These small models are actually worse than GPT 3.5 at some tasks (like recalling facts). You can definitely make models smarter at specific tasks (like tool calling, coding) but you can't compress the entire human knowledge into a 30GB file. It's just not enough bits.
Ideal local model would not know stuff like who Britney Spears is, best to leave precious weights for something useful.
Of course the line is very blurry but I'd be perfectly happy with local model that doesn't know anything about history, geopolitics, art or even biology etc. just coding, operating systems etc.
At one point there were specific -Coder release of Qwen e.g. 2.5 but they dropped that, still wondering how much better 3.6-Coder or 3.8-Coder would be when they ignore everything else
Models cant make any decisions if they have 0 idea that the feature exist in this language or in some general fact.
For example you would tell a model hey, become an expert in this language for me, search it online, it would still need to learn it and download the data to it's context and then increasing the memory usage, there's no way around it.
In my experience, the facts that are compressed away in small models are ones you don’t need them to memorize. They need familiarity with the essential concepts in a field, so that they will have better comprehension of material put into the context (and make better retrieval decisions). They don’t need to know very particular details which are a hallucination risk and should be derived from the context instead.
"Benchmark is stupid" and "model beats model on benchmark" are two different things, though. The second one is objectively true regardless of your views on the first one, right? To expect everyone to share your opinion that benchmarks are stupid is pretty weird, and just saying "no" to an objective truth is the definition of delusion.
But that's the problem with that logic. The author subjectively claims benchmarks are stupid (a debatable opinion) and treats that as inherently true while making the second claim (hence, you shouldn't say model A beat model B).
I can say I find the law stupid, so no one should say person A beat person B in court. But I did not prove the law is stupid; I merely thought it subjectively and demanded others to follow the second part because I believe the first part is true.
Saying that "if the law is useless, court cases are useless" is objectively true and cannot be argued with. But you still need to prove why the law is useless, not why you think it is and even then if people disagree and use the law as a reference, then it's not objectively useless and court cases are not useless as well.
For high concurrency, using the blackwell's native native W4A4 MLP compute path, nvfp4 is something like a 1.2-1.5x performance increase over FP8. We're doing data enrichment (so, tasks completed successfully + tokens/second) so the performance bump shows up in the tasks/month number.
I am just now getting the benchmarks running against 3.8 27b but I expect similar results from benching 3.6 27b at the same quant.
Unsloth usually also fixes the models when they bork something, which always happens. For Gemma for example the tool calling wasn't working for the longest time.
The talk around KL divergence is oversold. People talk about it as if it’s not a benchmark, but at its core it is in the same neighborhood! You get a different KL divergence number on different “calibration datasets”, so in other words it’s data-dependent. It is NOT a universal guarantee about the fundamental divergence of a model.
KL divergence is nothing close to a replacement for benchmarks. As flawed as benchmarks are, KL divergence is a barely useful signal. The fact that Unsloth only just started publishing KL divergences shows how unserious the quantization space is.
You're very right about KL divergence. I spent a couple days playing with the Gemma 4 models. That's 10 separate models (varying weights, MoE, QAT or not, etc...) with identical tokenizers. I treated 31B at BF16 as the gold standard, feeding Wikipedia snippets, and anthropomorphizing a bit:
Gemma 4 31B: "Um, if I really said all of that, I guess I'd say this next"
Gemma 4 26B: "Dude, I would've said completely different stuff" (large divergence)
Gemma 4 12B: "Umm, there's zero chance I would've said some of this" (INFINITE divergence)
Gemma 4 E4B and E2B: "Derp derp, I'm happy to say almost anything" (lowest divergence)
For models which are chat trained, they simply would not recite Wikipedia, so the divergence is almost meaningless. I thought about capturing a realistic coding session and trying to use that as the corpus, but you need to preserve the turn-based tokens and such, so I moved on to other things.
> The fact that Unsloth only just started publishing KL divergences shows how unserious the quantization space is.
Just wanted to say that this is a very important point that I totally agree with. People are obsessed with KL divergence, but it is yet to be demonstrated to be a descent proxy for agentic coding benchmarks.
It's a cheap to evaluate proxy for totally broken or not, which is a good start.
It also has a lot of resolution and not a lot of noise. Better would be multi-turn benchmarks with tools but getting good precision and accuracy for that is hard and computationally expensive.
When I tested various eval benchmarks on Qwen3.5/3.6 27B with Unsloth's quants, the scores usually dropped 0-5% between UD-Q6 and UD-Q3 depending on the eval.
Oh, goody goody goody! Can't wait for a Q4M with *MTP* support. Does anyone know if there are plans for it?
I am currently using Qwen 3.6 on RTX 3090 and I have to admit that without MTP it would be too slow to be acceptable for me (30-35 tok/sec without MTP, 60-70 with MTP). Without MTP I would just use OpenRouter and rather pay for speed despite having a capable local setup.
That kind of result makes me suspicious of benchmaxxing. Qwen 27B is 100x smaller than Opus 4.7. Is it really 100x more parameter-efficient? Two orders of magnitude is hard to believe. I don't have the hardware to run a 27B, but I'm curious what real world use is like. Maybe I'll have to buy some usage on a cloud provider to run my own tests, but this seems fishy to me.
Qwen small models are heavily coding focused, whereas Opus is everything to everybody (even if code is their bread and butter). The downside is they'll frequently hallucinate world knowledge so they need to be RL'd to double check their knowledge against sources and verify facts/library names/etc.
It's very agentic coding focused; and I'd say a good executor but certainly not Opus in scale; overall knowledge; long-horizon work and recovery; etc.
e.g. If you try to chat to it about something philosophical for example, or maybe a debate / creative writing, then you'll very quickly see how it is still a much smaller model at the end of the day.
Still, it's such a relatively accessible model to run, and I find a big part of leveraging smaller models is to give it well-scoped tasks; not too high level or ambitious ones. Very impressive for its size and the ability to run locally :)
In the local LLM communities there is a lot of respect for the Qwen models, but everyone comes to acknowledge that they do a lot of benchmaxxing after using them. Even at full precision they're never as good as models with similar benchmarks.
Any tips on the best approach at running this at an M4 Max 128GB? Token throughput was a bit slow with the last 27B one (MLX), ended up using the A3B variant but if I could get this one to reasonable speed I'd much prefer it.
I have an M4 Max (unfortunately 64GB). I have been running the Qwen 35B A3B one for a while now, after testing and benchmarking a number of models. That one was consistently the best in class for tasks like despamming, E-mail classification, OCR and dictation post-processing. It was also really fast (90 tokens/s).
I'm benchmarking the 3.8 model now, it seems it is better (near-perfect score on my E-mail spam filtering benchmark, best of any model I tested, ever). But it is slow.
One thing I would recommend is keeping an eye on MTP parameters. I tested and benchmarked extensively, and I use `--spec-draft-n-max 2` with llama.cpp. Longer sequences actually decrease overall performance.
As for running, I ended up using llama.cpp and its llama-server, with a bunch of scripts written by AI, because I got tired of LM Studio not implementing the image-related parameters which made gemma4 useless for OCR.
Go for a slightly more quantised version, and experiment with different MTP settings. I find that MLX versions are marginally faster on my 64GB M1 Max, but I usually use Unsloth's GGUFs via llama.cpp as there's a much greater range of quants available and I prefer llama.cpp. MTP sometimes also helps a little, but I suspect it's less helpful on my system than others.
This is the way if you need speed. It costs a little bit in smarts, but compare the MTPLX option listed above with the oQ4e-mtp quant using oMLX. The good cacheing layer in oMLX will help things feel faster for some classes of tasks in my experience.
Wait for the MTP variants that will likely be out within days. I'm on a 128GB Strix Halo box and for 3.6-27B 8bits I was getting about 9tok/sec (not great). With MTP that gets closer to 18 tok/sec (kind'a usable).
The MTP is available, but I'm definitely not seeing 18 t/s on the Strix Halo from the 8-bit quantization, even with MTP (more like ~10 with full context on long tasks). This is a slow model (but so was 3.6). What's your exact llama-server command that gets 18 t/s?
27B is a dense model so it will be slower with an MoE (A3B), but should have better quality? I still haven’t found very good uses cases on my M3 Max for dense models. Even if you can find a MTP version, it doesn’t help much, especially if you compare against an MoE with MTP as well.
Unfortunately, that chip just doesn't really have the memory bandwidth to run this (or nearly any) model at acceptable speeds. I have the exact same chip (M4 Max 128GB) and I've been trying to optimize a completely purpose-built implementation with Fable and this is just not possible. Even if you could reach the full 576GB/s, it's just physically impossible to exceed these numbers with the model's architecture:
2 bpw - ~85.7t/s
3 bpw - ~58.0t/s
4 bpw - ~43.9t/s
6 bpw - ~29.5t/s
8 bpw - ~22.2t/s
16 bpw - ~11.2t/s
without cheating. You'd have to exclude layers, skip operations, etc. basically do stuff the model wasn't trained for. And speed collapses so fast with context that even 2 bpw would be looking at ~37.6t/s after just 128K tokens.
MTP only improves the situation by up to 2x in the ideal case, while drastically reducing the performance floor. While optimizing a 9B model on this hardware, I've found that the GPU just doesn't have enough FLOPS to handle speculating more than one or two tokens ahead on a single stream, regardless of quant level, simply because of the arithmetic cost of the forward pass. The 27B model would be even more expensive than that, potentially such that it's already bottlenecked by the GPU itself rather than memory.
I wouldn't get my hopes up for the 35B-A3B either. Not only is it reportedly much less intelligent, but I hit a similar ~85t/s wall in practice (again with highly specialized inference).
Without speculation I can reach around 120t/s on Qwen3.5-9B and with n-gram speculation (not even MTP; this derivative didn't come with one) around about 150t/s on average. This is on the very very edge of what I'd consider acceptable for me to even consider using such a compact model. YMMV due to the silicon lottery but the situation isn't good.
I consider 'acceptable speed' to be around 150t/s. Why? Well, this is generally what it takes to keep me engaged with the output, rather than immediately switching to other tasks and checking back later. When I check back later, I have more catch-up to do at once, and I haven't been following the process. So I have to recall it, familiarize myself with the new progress, and sort of get back into focus with it, which is a lot of mental work (even if it happens quickly in real-time). I prefer not to have to do this because of how much work it is, so I prefer to watch the agent in real-time and try to follow its reasoning. That also lets me interrupt it quickly when I see it about to make a mistake, or see an important detail I left out.
The file "Just loads" on llama.cpp, the Unsloth https://huggingface.co/unsloth/Qwen3.8-27B-GGUF is an MTP file, I see mostly the same speed on pp and generation.
There has to be something wrong with those benchmarks, I find extremely hard to believe a 27B model can work similar or exceed opus 4.6.
Worth distinguishing knowledge/task benchmarks from IF / agentic. It doesn't seem out of the question that you can have a small model thats generally good at instruction following and long-horizon agentic, as usually in those cases any requisite knowledge is in the context.
Most of the benchmark improvements afaict are in agentic and instruction following benchmarks.
I think you've been drinking the "LLMs only improve by adding parameter counts" that SOTA labs are selling VCs to build data centers so they can keep eating through cash to their own benefits.
To the countrary, the reason Chinese models are excelling in the smaller area is because there's tons of fat in closed source models because of the crazy cash being thrown around.
There absolutely is space to improve intelligence and capabilities without lathering on more and more parameters.
For the full weights, unoptimized on vLLM with 2 Nvidia 6000 RTX 48GBs connected by NVLink, i only get 14 tokens/sec with open-code. For batched operations, it climbs to 55 tokens/sec.
For FP8, on a single Nvidia 6000 RTX 48GB, i get 13 tokens/sec on a single GPU and 46 tokens/sec batched.
The $1500 Intel B70 with 32GB of VRAM can run this model at max context with good performance, btw. If you don't want to drop $5-10k for running DeepSeek this is your best budget option for local refactor/small scale dev help
I understand the B70 is a bargain vs AMD and especially nVidia offerings, but to me it feels like I would be buying something that would feel too limited in less than a year. 48G would be much more confortable.
And I know the 96G nVidia cards are selling for over 10k$.
You can buy two b60s for $1300 right now (650 each) if you want a total of 48gb. Intel recently raised the price on all of their gpu's except the b60 series, so they are currently the best deal per gb I think.
32GB is perfect for models around 30B parameters. Since qwen has really hit the spot with their 27B dense models, I think it's a good bet. Also, 32GB is enough for other tasks such as image/video generation and loading multiple smaller specialized models
Sorry, I wrote autocompletion by force of habit. I simply meant it can complete code you have already created a structure for, which personally is very nice
I thought so, but thanks for the clarification. I am a little bit disappointed that local autocompletion models have been left by the wayside in favor of models post-trained for agentic coding. Both Codestral and Qwen-2.5-coder are more than a year old at this point, but local auto-complete seems to me to be such a great usecase.
The latest Qwen models (including 3.8 27B) do still support FIM-style in-editor code auto-completion if that's what you're looking for. I wouldn't want to use a large dense model like 27B for such a task (since FIM-style auto-completion really works best with low-latency responses), but it works.
Qwen3.6-27B has been the main LLM powering my little agentic stack. I have adopted the test and verify approach to any models allowed to run on my machine. When the "heretic" version drops, I will fire up the harness and test. Super excited to see how it stacks up against Qwen3.6!!!
If the benchmarks are a real indication, we now have a local model that is runnable on a high-end personal PC that trades blows with the leading model Claude Opus 4.6 Max from half a year ago.
It's worth running, even quantized. I liked Meta Muse Glimmer's outputs, but qwen3.8-27b@q4_k_s kinda seems way better. Haiku/Sonnet kinda pairing in workflows?
can you tell me ideas of usecases of 9 or 10B language models ? I cant find any usecases other than training a lora on them to give good bash commands for example
I use Gemma 4 12B in the 4-bit quantization for all sorts of vision tasks (image sorting, classification, description). It's also good for the same sorts of things for text (but there are probably better/faster models for text, 12B just happens to excel at vision tasks). The Qwen 9B is also very good for those tasks. If you need to do any kind of "search the web, grab some data, do some kind of action" tasks, these small models are perfect for that. Scraping data in a fuzzy format into a database or report or spreadsheet, producing a dashboard of news, etc.
Small local models can also be used for sub-agent tasks in most agent harnesses. But I'd probably run a larger MoE for that; they're faster and have broader knowledge. The dense models, even very small ones, are not blazing fast.
I don't code with any models small enough to run locally, at least not so far. Qwen 3.8 27B might be the tipping point, though. It's looking really promising, though it's probably slow enough that I won't ever actually use it. I'd rather pay $100/month for a faster model, even if Qwen 3.8 turns out to be smart enough for most of my work. Running it locally with the 8-bit quantization is going at 12-30 t/s, depending on how much context it's chewing on. So, if all you do with AI is coding, then you're better off doing it in the cloud.
But, there's lots of things a small model can do that aren't coding.
Yeah this is the first model I have been able to run locally that actually feels useful, this is unreal I am considering cancelling my claude sub and going to just api (maybe GLM?) for really hard tasks.
Yeah I was thinking open router but I will look around at options, I genuinely think this model is good enough for like 90+% of my use cases, and the top frontier models are still not that good at architecture so I have to do that myself still so I won't be losing out.
Do we really feel like it’s the governments job to regulate OPEN source AI.
At rely health, we use OSS models in a HIPAA complaint and SOC 2 complaint environment to take advantage of asymptotically $0 intelligence to provide best in class care navigation . This should be industry standard -
I wonder if Anthropic and OpenAI possibly missed the window to go public. A 27B open-weight model trading blows with the SOTA from just half a year ago is not great news for trillion-dollar investments...
Trading blows in some benchmarks is a bit exaggerated. If you try the model, `xhigh` is basically feels like the `max` mode (i.e. massive thinker and extremely presistent), and the amount of world knowledge and intent understanding is nowhere close to an Opus class model even from 6mo ago.
It's still very useful, and it'll probably displace a good bit of API spend; but it's not really "trading blows with SOTA from just a half year ago". A bit overblown on the Anthropic/OpenAI has missed the window I reckon.
Also, on the open weight frontier side, Kimi K3 is pretty expensive, and Deepseek V4 Pro/Flash is getting a little less juicy with price increases.
Lacking world knowledge is fine for me, I rarely want to rely on the model’s training anyway when there are plenty of great search options to integrate with.
IMO AI companies are selling the infrastructure and easier access to the models, beyond competing to get the smarter model. We can run Qwen locally, but at what cost? A good enough GPU for running 27B is more than 4,000 USD where I live (Brazil). Even if I were to spend 100s of hours in credits I would never reach a point where the 4000 USD investment pays off. Unless I'm doing it for the sport, building my own LLM rig is not financially viable.
This is without saying that company clients of Anthropic or OpenAI will not have the upfront money to get a RTX 3090 for each and every employee. Much better to sign a deal with the LLM providers and get cheaper token price with a sales person.
That’s quite inefficient. Get 2x RTX 6000 Pro and run DeepSeek v4 Flash for the whole company. This may or may not be cost effective compared to cheap third party providers, but it lets you own your own data, is not actually that expensive, and reports suggest that you can get 2k tok/s or so from a setup like this (at respectable batch size, but that’s exactly what you want in this situation).
I’m not sure I believe that the unquantized performance is as good as people are claiming, but even light quantization would make it fit easily.
Another way of rephrasing that though is that the public (including large pension funds, retirees, etc.) just dodged a bullet and aren't left holding the bag.
Nevertheless, I doubt anyone's going to emerge unscathed when the valuations start falling though. Vast sections of the American stock market will be wiped out yet and with it the global economy.
Why do you assume 2-3 American companies failing will have a more negative impact on the economy than positive impact of the widespread availability of self-hostable, open-sourced, and efficient models?
There will surely be tons of new companies that make a business off of hosting these models and even tuning them for specific purposes. Not to mention the massive pricing difference will benefit people who actually utilize them
Firstly, I don't think it will be just 2-3 companies that fail. Due the circular investing between companies in the AI ecosystem, the fallout is going to be much bigger.
Secondly, even with that being true, I think open models are a great net benefit for all mankind in the long run.
I think both of those things are true at the same time.
Even Zuckerberg isn't as deluded to think that there's money in constantly releasing closed models and hypserscaling their deployment, which is why we got muse-glimmer 30b last week. I have been playing with it on a 32gb MBP w/ M2 Pro. If prosumer-grade hardware from almost three years ago can now offer a novice a way to host a possibly serviceable software development agent, then there's little incentive to pay out for LLMs like Anthropic's or OpenAI's over the long run.
For those commenting on the long reasoning, it may be interesting to know that the reasoning effort is set to xhigh by default [0]. Other possible values are medium, low and none. Flag for changing it in llama.cpp below, but note that the long reasoning seems to contribute a great deal to the quality.
Architecture thread! Afaict they continue to use gated attention + delta net, which was also adopted+adapted by K3, but im surprised theres no improvements to the residual stream (deepseek are using manifold hyper-connections, kimi have attention residuals) ?
Model benchmarks are useful, to a point, but it is the long tail of things you do with the model that determines if it's good at a wide range of activities. Ant/OAI, to their credit, build their models -- even the small ones -- so they follow instructions and do tool calling well, without the system prompts confusing them. This is especially important for long-horizon tool calling.
So one open weight model might "meet" Opus or whatever on benchmarks, but then fail to follow a simple answer format and also tool call correctly. The models are whipped to within an inch of their lives to strictly adhere to their post training quality gates.
Apart from model performance, what harness are people using to come close to Claude Code or Codex workflow styles with tool use, conversations, loops, remote control, etc.?
Try manually asking both more discrete esoteric knowledge questions. Or use benchmarks which are less coding focused. The 3.6-35B-A3B with post-training may do well in coding type benchmarks and math but the density of its knowledge falls off in my experience (vs 3.6 27B dense Q8-K-XL unsloth GGUF) when you need to use it for less commonly used domains of knowledge.
My use cases try to avoid accessing world knowledge in the model (I give it access to web search for some adhoc RAG), and ya, I'm just focused on coding so that's the only place I'm looking at right now.
I think you may find that the dense 27B also does better if challenged with more rare coding tasks, less common or weird languages or things that aren't well represented in the active 3B parameters of the MoE model (eg: NOT css, javascript, python, c++, etc).
True. Benchmarks have led me astray before, I'll have to actually benchmark it directly in Goose with my real use cases right now (which are mostly writing python code, so it probably still doesn't apply?). Still, I've never noticed a benefit to using a dense model rather than an MoE for coding with an agent tasks.
Better than Opus 4.6 at computer use? Comparable with it for SWE? Am I reading this right?
I’ve heard rumours about AI shops optimizing for benchmarks. I also don’t think Qwen/Alibaba would be crazy enough to claim something unless there is some truth in it. Would love to see a side-by-side with Opus 4.6 on categories where Qwen 3.8 27B aces it.
Is there any way to turn off thinking if I'm using Ollama? In my particular case, the Ollama API (the software I want no-think for is tied to Ollama's bespoke API). If not, I'll stick to 3.6 for the time being...
I think I am going to buy a second rtx 3060, as 27B has been just outside of my range for to long, and this looks like the parameter count tipping point
whats the tok/s you get on that. I have heard a few claims of around 30-50 with mtp, but for how cheap the setup is I am surprised I don't hear more about 3060 stacks so I assume there has to be some catch.
I used GPT-5.6 Sol high to optimize it, and it claimed it was getting 50. I'm seeing ~40 on my goto smoketest: "Make me a vector add in CUDA".
Funny side note. It successfully one shot the program, but it wasn't able to run it because there literally wasn't enough VRAM left to allocate CUDA memory. Watching it try to debug that was fascinating. I'm pretty sure it would have killed the llama-server (and thus itself) if it hadn't been running in a separate container.
I've been experimenting with a few settings in my 4090 , and if 3.6 run at 90-110 tps, 3.8 staya below 80 tps and it's most often at 60 tps.
I'm using flash attention, mtp speculative decoding (n=2). I've looked at the club-3090 repo, but haven't found anything meaningful but get back to 3.6 performance
Couldn’t afford claude pro so I built an web based DSA coach that coaches you on DSA and System design in a socratic way. It uses a qwen 1.5B coder model and inferencing is all done on a CPU. ( who needs a GPU anyway )
Yeah, I would appreciate if someone could make sense of the pricing differences between these models. How can a provider run DSv4F at lower cost than a 27B dense or 35B A3B model?
Does it come down to utilization and/or specific model tricks and efficiencies (attention, kv cache, etc.)?
> and more memory/computationally expensive to run any way I look at it.
Just speculating, but look at attention. 35B-A3B context is heavier -- about 20GB per 1M tokens vs ~7GB. So if you have ≥10M tokens of context per inference node, then DS4 Flash uses less memory than Qwen 35B.
Qwen attention is probably more computationaly expensive as well.
Also I guess not much interest/competition in serving Qwen edge model.
Yeah, it should be basically free. No idea why it is not. I guess KV cache taking up RAM and possibly bad business sense or amortized engineering costs, I honestly do not know.
The default reasoning is set to 'xhigh', if you want to compare with the past or reduce the time (if you can take the hit in output quality) then you can pick 'high' or 'medium' as well.
I wish this can run directly on my RTX 4090, seems like 30B is the sweet spot for dense model to run locally, sadly RTX 5090 is very expensive and I need a new PC and new power supply(and UPS) to run that, adding a second RTX 4090 is another option, but not sure if my PC can do that yet.
even a 3090 will give you the VRAM headroom. i run Q8 on an 3090/A6500 combo. well, Q8 of 3.6-27B. I'm building the Q8 GGUF for 3.8 now, assuming mine will finish before someone else's.
Can anyone who has that specific personal test he tries on different models , and tries this model , to tell us here if possible , how good or bad is this new model ? compared to others ?
running an untouched, vanilla 4-bit version (Q4_0) I baked myself today (benched it against Q4_K_M (16gb) and IQ3_M (12gb), Q4_0 (15gb) is king)...
this model--
1: over 60% faster than qwen 3.6 version of the same dense 27b model, same engine setup (don't ask how, im not sure either)
2: has better reasoning quality, less "loopy" with its thinking patterns.. most certainly the smartest model on my roster currently
3: has the longest task horizon ive ever experienced (locally or otherwise)...I sent it a bunch of compressed ideas for an app, it sent me back the largest python app ive ever seen in one single ai response pass (80kb text file)
thanks qwen!! hoping to see the full model range get released...
There is a down to earth guy on YT that performs a series of tests against LLMs running on non-god-tier commodity hardware. He will likely be testing this soon enough.
So much potential for that channel. He's got a nice range of tests and a no nonsense presentation style.
However, watching tests of heavily quantized models that weren't designed for it (non-QAT) is frustrating. There's no way to tell if the actual model fails because it's dumb or if the lobotomy made it that way.
I use the Qwens as a vision model for my DeepSeek V4 Flashes to handle. But the Qwens run on old RTX A6000 Ampere. Does anyone know if there's any news about INT4/AWQ quants for the RTX A6000?
Same, with a multimodal DSv4 Flash I would just stop paying attention to things. Very smart, and at 260 tok/s it's too fast to care about anything else. If you ever graft something like that I would love to hear about it.
Yes, I have a very dumb flow. The harness has a describe_image tool that takes an image and a prompt and so DSv4 Flash uses it to get an idea of what it's looking at.
Yeah, I might just replicate what you're doing. Main issue right now is just finding spare vram to actually run another model in parallel... And yeah, if I train up a vision adapter somehow, I'll try to put it up/post about it, seems like we're getting the killer apps for local LLMs right now, where it's just feasible enough if you're enthusiastic enough to be a bit economically irrational, and just useful enough to sort of rationalize.
The Hugging Face page (https://huggingface.co/Qwen/Qwen3.8-27B) shows that Featherless supports the model (and includes the inference widget to try it out directly on the page)
strong visual reasoning apparently, which is nice. still lacking native audio however. hoping for more companies to embrace the spirit of something like `gemma-4-12b-qat` for actual multi-modality (text, image, video, audio).
Why don't qwen/alibaba host the model themselves? I was looking forward to trying it on their coding plan. Google are the same way with their Gemma models.
I thought the same. But why claim something so shocking when it can easily be discredited and puts your reputation at risk? If they’re claiming Opus 4.6 level, I expect it to at least match Sonnet 4.6.
Since Qwen 3.6 27b outperforms Gemma 4 26b in most benchmarks I'm not sure the value - also Gemma 26b is a MOE model whereas this is a dense model, so not typically direct competitors at their sizes - Gemma 4 31b comparison would be interesting though.
Is it just me or is 3.6 27B Q8 K XL (Unsloth) holding up better in sustained token/s rate as the context fill increases over time? The token/s rate seems to be much higher for a time period deeper into context than previously seen.
At least as compared to 3.6 27B in the same quantization.
I don't know much about the production of these models. How hard would it be to 'fork' something like this and have it not be full of CCP indoctrination?
What a weird mechanism. If someone is judging a thread/topic/event impact by the number of points it got, then doing this unfairly degrades that thread.
It should have deduped by user and combined the 168(at the time of writing this comment) + 289 points. Just add the twitter link from the previous thread as an additional link in the description, like you normally do, move all the points over, and remove the old thread.
I'm hoping to see folks distill this with current generation Opus / Fable reasoning traces. I have had my best results locally so far from Qwopus (Qwen 3.6-27B w/ Opus 4.6 reasoning distilled). This looks GREAT and I am definitely setting this up later today.
I found this kind of amusing while running it (using Pi as the harness). Don't know if this is evidence of intense fine tuning from Claude but it smells like it...
" The user wants me to explore the repository at XXXX and report back. Let me start by understanding the project structure, reading the
CLAUDE.md file, and getting a general overview of what this repository is.
Let me start by reading the main project documentation and exploring the directory structure.
I'll take a look around this repo. Let me start by getting a lay of the land.
read resource CLAUDE.md (ctrl+o to expand)
ENOENT: no such file or directory, access 'XXXX/CLAUDE.md'"
Why is anyone even using video cards these days? You may as well be burning cash.
This is the perfect candidate for just splattering it on your nvme and then reading it off there and into memory. All of these run perfectly fine on simple m4 silicone:
Nvidia and AMD both have their own unified memory laptop SOCs, now. Apple Silicon's GPU is relatively weak, it's one of the less-efficient ways to use 100w for compute.
Even the fastest Apple Silicon chips like the M5 Max and the M3 Ultra still put up worse GPU compute performance than last-gen laptop RTX 4080 chips. And they don't scale, the largest M3 Ultra cluster you can configure is still ~2,000x smaller than a DGX SuperPOD. There's a reason Apple discontinued their rackmount hardware, there's very little demand for Apple Silicon in the datacenter.
SWE: model is strong for its size. It one-shotted the Web app, had no bug. The Rust rewrite only had one bug (reordering didn't work immediately - fixed in one prompt). Committed locally then pushed to my GitHub (https://github.com/DexterLagan/RusTODO). Can't complain. If it can do that reliably, I can use it to make whatever I need.
General knowledge: I usually ask 2 questions many small models get wrong: summarize Operation Trojan Horse by John Keel, give publication year. Summarize the Ariel school incident of 1994. Qwen3.8 got the first question right, along correct publication year, gave glorious detail on Keel's theory, but got the second wrong. It thought that school was located in the USA. Ah well.
Performance on an admittedly overpowered laptop: 15 tokens/s in power save mode on this MacBook M5 Max 48GB, and 30 in performance mode. Perfectly usable for local coding through OpenCode.
Verdict: very nice local and free backup to my usual GPT/Claude/DeepSeek for code. Good for Web searches via Brave search tool calls. What more do you want from a small local model?
I'm not sure that this means anything. You're asking a ~27GB file to have losslessly compressed the entire training set (which apparently is a large chunk of the entire internet). That's not possible. Whether it happened to encode these particularly obscure facts losslessly or vaguely isn't really telling you anything about how good a model it is.
Qwen3.8 also seems to know French quite a bit better than Copilot, at least on common expressions. I have yet to run more tests for languages, but I'm baffled by its finer accuracy on the few sentences I asked it to translate, which were botched by Copilot. Go figure.
Gemma 4 reasoned through it more implicitly, while Qwen 3.8 reasoned more explicitly. Laguna and Muse Glimmer failed hard on it, though they're useful for other tasks.
The VRAM usage seems way less efficient than Gemma 4 or Glimmer though, with 32K of context taking 2.5GB of VRAM. With those, even with MTP or a DFlash model loaded, you could still fit 256k-768k of context. With Qwen 3.8 27B I can't even fit 128k if I quantize V to Q4_0. Maybe with some trial and error I can find some settings that perform well enough with a larger context window that it's still useful for longer tasks.
Lots more testing to do, though I was getting some decent results out of Muse Glimmer which was more than twice as fast and supported huge context windows, managing to solve some bugs that Gemma 4 struggled with. I can't even begin to throw that task at Qwen, because just the prompt alone would use the entire context window and then it would reason for probably that same amount.
If you've got a 32GB card, it should be a decent model even if it really is memory hungry.
EDIT: Tried a few kv cache quantization settings, but it failed with those. I designed this benchmark to be pretty brutal in the face of KLD and any reasoning quality loss, so it's not too surprising. Gemma 4's QAT held up pretty well, at least and could consistently complete it.
What quantizations and context lengths did you use for Gemma 4 and Qwen 3.8 27B?
I am asking because I can't even load Gemma 4 31B on my GPU with any reasonable quantization (even with small context), while I can run Qwen 3.8 27B with large context and good quantization...
Gemma 4 26BA4B would get close, but not quite and sometimes even get stuck in loops despite a repeat penalty.
Do not use any newer updated templates or Unsloth fixes. Use older official templates that released with the models on the huggingface repo. The template here worked: https://huggingface.co/google/gemma-4-12B-it/tree/657684fef0...
llama-server --model "model.gguf" -fa on -np 1 --jinja --ctx-size 262144 -b 768 -ub 768 --cache-type-k f16 --cache-type-v q4_0 --repeat-penalty 1.1 --chat-template-file "chat_template.jinja"
If you don't explicitly point to the template file, then llama.cpp will either use the template inside the model file or it will use its own template copy and your results may vary. Obviously some of the template fixes are useful to people, so it depends if you're having problems with tool calling or can't fix the tool calling in other ways for your scenario.
My experience with the QAT models was that quantizing v to q4_0 gave me better results than q8_0 or even f16. I think the Gemma QAT models may have been QAT trained to expect a q4_0 quantized v cache. If you're not using a QAT model, I would leave both at f16.
Another thing aside from using the QAT models and a Q4_0 v cache since you're having trouble fitting the models, is that you don't have to use the mmproj if you don't intend to use vision. If you need vision, but are hurting on VRAM, then you should be using --no-mmproj-offload. That will keep the mmproj loaded in system RAM instead of on your GPU. Loading images will be a little bit slower, but it can still be quite fast and you'll have more breathing room on your GPU. If you don't provide the mmproj file on the command line at all, then it won't load it anyway. If you're using some program like LM Studio, a simple thing you can do is move the mmproj and mtp files out of the directory for the model so LM studio can't find them and then it won't load them at all.
For Qwen 3.8 27B, doing any quantizing definitely hurt results a lot, so in my case I used: llama-server --model "Qwen3.8-27B-UD-Q4_K_XL.gguf" --spec-type draft-mtp --spec-draft-p-min 0.35 --spec-draft-n-max 2 -fa on -np 1 --jinja --ctx-size 65536 -b 768 -ub 768 --cache-type-k f16 --cache-type-v f16
Actually perf (speed) is going to mostly on token output, and here Qwen 3.x historically tends to lose badly as it tends to overthink a lot. I'll be running evals on 3.8 myself this weekend to see how its reasoning levels perform.
I assume that AA will have 3.8 numbers soon and Intelligence Index vs Output Tokesn per Intelligence Index Task is a decent way to view that: https://artificialanalysis.ai/models/muse-glimmer?intelligen...
Qwen might be useful to bring out for a second opinion on some more focused details that are largely information complete. Like, use Glimmer to bring together all the relevant critical data and evaluate what the actual problems are, then maybe prototype a solution. If it's still acting up, maybe throw the resulting context at Qwen and let it meditate on it.
I think there was some study done where ideally you would want to throw a bunch of different models at a problem since they don't all have the same perspective or diagnosis on what the problems or the solutions are.
From my testing so far, Qwen 3.8 is better at code but it tends to meander and take forever if it has to look in a lot of places. Glimmer will use like ~1k tokens to formulate a plan and Qwen 3.8 will routinely go over 10k
For some usages that speedup more than makes up for it being inferior to Qwen intelligence wise.
Maybe it's implicit that you're using llama.cpp (although you don't mention GGUF), but it's hard to reach concrete conclusions about the model architecture based on one implementation in one runtime.
Llama.cpp checkpoints and moves snapshots of the cache to main RAM for faster resumption after truncation.
I don't expect you to blab publicly about your private benchmark, but what sorts of reasoning does it require?
#1: it does not require deep world knowledge, because that's not what local models are for.
#2: it directly attacks drive-by understanding, overly linear processing training, poor attention mechanisms, poor reasoning patterns or lazy assumptions that ignore very easy low hanging fruit.
#3: it requires solid instruction following in the face of errors. a lot of models will run into errors and then fall back into some kind of error recovery process that bypasses instruction following.
#4: does not require prompt fine tuning to tweak to each individual model. they all seem to understand.
#5: not unfair. almost every model demonstrates in their reasoning that they have the necessary information that if reasoned about appropriately, could arrive at the correct answer.
#6: not designed to add unnecessary complication that it is intended to exhaust reasoning budgets of any sort, so it is not inherently unfair to models that reason a little more or less. for example, it does not require unnecessary reasoning soaks (ie: hiding the prompt inside base-64 encoding)
#7: has real world use and is probably applicable to overall ability to generalize.
#8: can be scaled up as models get better.
#9: is a very good indicator of how bad a model is falling apart under various inference settings.
Want to say more about these private benchmarks? :)
Bicycle is the right shape. Pelican beak is excellent. Nice background. Most importantly, the pelican has one leg on each side of the bicycle - that's very rare.
(No chain on this bicycle though - in the reasoning trace it says "already chainstay... skip chain detail; maybe a small chainring.")
I ran that on an M5 Max MacBook Pro using LM Studio and their 17GB GGUF: https://lmstudio.ai/models/qwen3.8
It took 21 minutes(!) and used 22,276 reasoning tokens to produce 3,223 tokens of output.
(For the "they're training on your benchmark now" crowd, all of that cheating didn't prevent it from spending 20 minutes thinking about the task first! You can see the reasoning trace in the link I shared.)
For comparison, here's one I got from qwen3.8-2.4t-a95b on OpenRouter, which is pleasingly animated: https://tools.simonwillison.net/markdown-svg-renderer#url=ht...
My SVG rendering software was stripping some attributes. Here's the Gemini 3.7 Flash pelicans in the fixed renderer: https://tools.simonwillison.net/markdown-svg-renderer.html#u...
It’s obviously a problem of fundamental understanding and demonstrates that reasoning is more “directionless rigour”.
Besides which, "fundamental understanding" isn't a binary. Humans can also have a deep understanding of a subject and still make errors. (I do, anyways.) I do not think "fundamental understanding" is itself well enough understood to say definitively that LLMs do or do not have it.
*Some models have multimodal capabilities but these are separate weights, I believe it's only engaged in processing images. From the reasoning trace, the model never rendered the result -- only instructed the user to do so. I welcome correction if I'm wrong, I only have a surface-level understanding here.
"pelican = ocean bird = nautical theme = this means I should put a captain's hat on the cartoon Pelican"
I think in some sense it understood the absurdity of a pelican riding a bicycle as intentionally nonsensical, so it went with a cartoon theme for the design and the shading. Worked out really well.
But, not animated like the Qwen 2.4T one you showed, so minus points for that
https://imgur.com/a/IgYKnmk
Took almost 90 minutes on my Threadripper Pro, and feels well worth it.
28k tokens from the Q8 model
Chains exist. Red scarf is proactively added. ("Maybe a scarf blowing in the wind for charm!) No hands/wings, though.
Generated 30.2k tokens in total and took 52 mins on M5 Pro in low power mode (it will possibly take less than half of that in auto energy mode).
This took 56 minutes! - but it didn't have exclusive use of LM Studio, I ran a few other test prompts while it was working.
It is also worth noting that it shaded the pelican case like it was a product photo taken in a studio, there's a virtual light source above and to the top left corner of it, then shading deeper color of orange as the body area progresses towards the lower right corner.
https://gist.github.com/hansale/ed9e73fe35165a58ea2af6b1632a...
Some will say conflating flying and flying a kite is bad, but I like it.
I switch off reasoning by default and for starters I ran 3.8 the same way as 3.6 on vllm on a DGX Spark box. Open web UI prompted.
My word! It looks rather decent. Much better than previous efforts.
It's up to you if that counts as benchmaxxed.
"Need be helpful concise", "Need maybe not overdo", "Need ask!" Almost caveman.
I have an (unsourced, vague) suspicion that this rather unique thinking trace pattern is actually hobbling the MTP predictions, which seem to perform poorly.
Other notes: it uses the trick of repeating the prompt in the thinking trace.
It also worries about hidden chain of thought appearing in the final answer. It talks about "desired oververbosity 9", which is new. A bit GPT-ish.
It is being extraordinarily thorough in thinking through one of my code requests, but I don't know if the net result will be any better than the 35B MoE.
I asked it to ask me clarifying questions — it did, and it offered me a list of defaults I could simply agree to.
I don't think it is necessarily overthinking in the looping sense, but it is in the being exhaustive sense. I need to explore how it does with a tighter reasoning budget.
I am impressed but I am definitely in Camp Please-35B-A3B-When? here, because on an M1 Max this isn't really practical. I hope they do one, though I think they may not.
All your examples align with this too. It just sounds to me like its optimising by removing the superfluous English words and thinking more Chinesely.
Frontier labs have already been doing this for a while, verified in smuggled traces from OAT/Ant.
Simply a way to reduce tokens.
Oh dear. I need to try to understand what is going on here.
ETA: a bit of testing before I climb the wooden hill to Bedfordshire.
LM Studio doesn't seem to display the little dropdown to set reasoning effort, so I bodged the chat template on load to get it to choose 'medium'.
As soon as you switch away from xhigh, it goes back to thinking in normal sentences like Qwen 3.6, rather than in sort of quasi caveman.
And you get all the Wait, Actually, No wait… stuff back.
And it is behaving a lot more like it used to. So that is pretty interesting.
Compared to closed weight (especially unreleased and access-limited) and open weight/source large sparse MoE LLMs/VLMs, open weight/source small dense models benefits public the most because they just reaches more people.
Compared to Qwen 3.6, 3.8's thinking style changed drastically. With xhigh budget, it thinks a lot MORE, and longer thinking session directly translates to better performance. This tradeoff between performance and computation, memory, etc. is meaningful to me.
However, because Qwen 3.6 and 3.8 share the same architecture, with 32GB vram, llama.cpp, IQ4_XS model, MTP and FP16 mmproj, I can only get 200k context, which is not good compared to 640k context of muse glimmer. Hopefully this problem will be solved in Qwen 4.0 release.
Edit: Absolutely blazing fast! Getting 163 tokens/sec on WSL and it generated a pretty sweet Pelican.
https://gist.github.com/hansale/ed9e73fe35165a58ea2af6b1632a...
[0] https://huggingface.co/froggeric/Qwen-Fixed-Chat-Templates
Like maybe I just misunderstand what's the hard part but wouldn't you assume that people who can put together an impressive model can also write a proper jinja chat template for it?
One really would think these companies (including Google) who spend many millions of dollars on compute could write a few hundred lines of Jinja correctly, so their investment works optimally or at all.
But they don't.
Then a couple of individuals on HuggingFace fix it, either a 2-person startup like Unsloth or a volunteer like froggeric.
I also don't understand how this repeatedly happens.
My favorite theory is that they had many model variants internally, each using a slightly different chat template, so when it comes to the release they are not even sure what to use any more.
I’m sure there is some basic testing but it might be agentic (LLM likes its own output) and maybe just some human smoke tests.
If you're using llamacpp, turn on top-n-sigma with sigma of 1, turn off top-p/top-k. You'll thank me later.
You might understand this as "The capital of France is..." and the model isn't always going to select "Paris". Sometimes it will start a descriptive sentence or even get the answer wrong.
That selection of the next token is what these settings control, and lots of sub-optimal selections compound over time to produce a junk response.
Top-P: example setting 0.9. Select tokens whose probably accumulates to this number. So say you have tokens with 0.7 then 0.2 then 0.1, the last will not be selected because the first two tokens already accumulated to >=0.9.
Min-P: example setting 0.05. Don't select tokens less probable than this value. So a token with 0.1 would be considered, a token with 0.01 would not.
The purpose of all of these is to exclude very unlikely next tokens.
Until we go advanced geothermal or we crack fusion, energy is dirty. Read my inference or link me to yours so we don't boil the planet.
I'm sure there's some benchmaxxing going on, and some things you get only with a a larger model.
But I'm feeling pretty confident if not by Gemma 5 than by mid 2028 we'll have local models that are almost always as good as Opus 4.6 was and in many cases far better.
[1] Without searching the internet. And even if you allow it, you'll get much worse results because search means browsing and parsing the top results, and search results are horrible, whereas internal knowledge from training encompasses the entire internet plus all books including very niche stuff.
Fortunately, that isn’t necessary! What LLMs need is a level of fluency with key concepts so that they can (1) make effective use of retrieval tools and (2) understand the material in the context window. 30GB-sized models can absolutely store enough knowledge to do this.
Here is an example from the field of law. Most lawyers who have litigated contract disputes in California know about Civil Code § 1717, which makes any contract providing for attorney fees to a prevailing party mutual, so that even if the contract was written to be one-sided, it won’t be enforced that way. It’s a simple enough concept, but there are many more particulars to it, such as what happens when the fee provision is only written to apply to part of the contract. (Answer: it depends on other facts.)
When a lawyer recognizes that they’re in a situation where § 1717 is relevant, the first thing they will do is pull the statute and read it, because nobody has it memorized. And they don’t need to.
A downside is that you can't just download a lot of that knowledge, vs with the weights the copyright infringement has been outsourced to the lab. Nor can you just search for the info because the internet as a whole is increasingly aggressive at blocking anything that looks like an AI agent.
I'd love to see more retrieval powered local AI-- I think it's an area that open source development could excel. ... but there are advantages of having the knowledge in the weights!
Perhaps what needs happen is for someone to make an "ultrapedia", an AI restatement of a huge library of reference works-- created expressly for the purpose of being a locally stored corpus for AI agents.
...I'm just asking questions here... how sure are we of this?
If you'd asked me six years ago whether we could compress all of human knowledge into a 1 TB file, I would have said no, and yet here we are.
If you can do 1 TB, why not 30 GB? It's, like, within one order of magnitude.
On a more serious note, it depends on your cutoff for "entire human knowledge". It's easy to prove for a generous interpretations of "entire human knowledge" that it can't be done, but hard for something like "all useful human knowledge".
Moby Dick, uncompressed, is ~1MB. Compressed, it's around 500KB.
I feel fairly certain that one could fit all of the textual knowledge required to cultivate a world-class <insert name of preferred professional knowledge worker> in <60,000 Moby Dicks. (Arguably in <5,000 Moby Dicks with intense effort/pruning).
Or maybe not, maybe there's a world model needed for human level at any profession that is very hard to quantify and requires more than 30GB by itself.
I just think about all the content I’ve consumed in my life to become a professional software developer and I would be very surprised if it couldn’t be adequately represented by <30GB of uncompressed text. Most of the work was in “training”, not really in data.
The “foundational overlap” of K-12 education is identical for most professions with 2-8 years of “finishing” on top.
My mental model is that the budget is spread across maybe 20% genetics (most of our instinctive/genetic information is surely pretty useless), 50% k-12 education, 30% for professionally-specific knowledge.
I think the problem is that the conversation drifted from "all useful human knowledge" to "enough knowledge to do my job".
Both a human and a current large model will bring up unprompted some tax regulation that applies in your district that could matter to your task. A small model won't know the tax code of every city in the world, as it's probably impossible to fit, and it's the kind of thing that unless you already know about, it's very hard to search for, let alone come up with unless specifically prompted about something tax related.
Unless we start having different small models trained for a certain profession for a certain city, I think we'll need those trillion parameters.
I’ve looked here: https://github.com/earendil-works/pi/blob/main/packages/codi...
I don't want to try to take my best educated guess at what the best design is BEFORE implementation - especially if you're designing a feature for a codebase you're not an expert in, you don't know like the back of your hand (i.e. one that is mostly or entirely LLM generated).
What sounds good on paper - often times becomes unideal in practice when you get to the reality of implementation.
It may not be worth re-architecting your entire system to get to a "pure" design that would be the best - all things considered.
Instead, I'd like the model to independently design many plausible and coherent good solutions, then implement each of them, then intelligently pick the few winners (after its fixed any bugs that could be causing promising solutions to look artificially bad) - unless there's an obvious one - and then give me the data I need to make an informed decision on which one to go with, all before I even look at the design or implementation.
You're not getting this from a one shot prompt from a 30B model today. You can't even really get it from Sol or Fable - IME. But you can get somewhat close.
Even Fable is bad at this, I would constantly have to fix it going down architectural dead ends or just making obvious mistakes.
Which sucks for people that want LLMs to do everything like a genie, but does mean senior engineers have a few more years before they become redundant.
Original images: https://image.non.io/neonRamenDesigns.webp
Qwen 3.8 build: https://html.non.io/neonRamenQwen3.8-27b
Overall I'm very impressed with how well this did. It's a big improvement over 3.6, and it feels on-par with some much, much larger models. I think this one is on-par with Gemini 3.7 Flash.
One thing to note - the build for this on my RTX 6000 pro blackwell took a long time. Easily one of the longest builds I've done. It took around 2 hours to build the site. Obviously we'll have some quants for this soon that will accelerate things, but I was still surprised with how long it took.
Comparison builds from this week:
https://html.non.io/neonRamenGemini3.7
https://html.non.io/neonRamenGLM5.3 (note: non-multimodal)
As capable as it is, it's hard to justify using it when a competing model (e.g. Gemma4:26b-a3b) can consistently achieve the same or similar response with only 1/10th as many 'thinking' tokens, achieve much higher tokens/second, and take a small fraction of the time. I suppose 'YMMV' depending on your use case.
Also, I haven't used it enough yet to see if it's prone to infinite looping, but its predecessors sure were.
In general it's kind of a benchmaxxing/distillation (I don't mean that pejoratively here, read on) artifact where test time compute's purpose is to refine/zero-in on a particular input:output pair. Basically they're trying to "remember"/re-derive the answer rather than arrive at it deductively - it gets comical/absurd when you see it expending a huge amount of tokens trying to figure out the answer to some trivial question or response to some input, as if it were a trick question or the choice of wording was of the utmost importance. But it's not a bad thing when the output matters or the task is hard: basically the model has been trained to respond/act like a much smarter model and it's probably better for it to overgeneralize that behavior.
Also, IME Gemma and most other local models that support thinking tend to have the same problem, and it's partly only a "problem" because they give you access to the thinking tokens themselves (which you don't in many cases see when working with frontier american models) and you can actually see how they're getting spent. For the local or bulk (runnin on owned or rented hardware) use case where you are paying in time/watts rather than purely by the token, IMO it's a good idea to just look at the task success rate / time per task and not worry about the raw thinking token count.
https://huggingface.co/froggeric/Qwen-Fixed-Chat-Templates
Granted, it's still much lower tokens/s than you'll get out of many MoE models.
Edit: Even set to medium or low there's still a lot of second guessing, less consistency, lower 'acceptable response' rate, and slower/more token churn vs gemma4:26b-a3b. I think gemma4 is just a better 'general purpose' model.
I don't usually rely on AI for much (I'm actually kind of anti-AI, although I follow stuff like this enthusiastically because of the rapid advancements, "average joe" access, and openness), however for what I asked? It was spot on.
The subject was a bit personal, so I won't share. I was just curious what AI would say about the situation and it definitely surprised me, especially since Qwen, while usually great on development/coding stuff, has shown weaknesses in other areas.
Definitely a solid release, and this one runs on my 4090 with minimal loss of quality!
Did I miss something, is it possible to have that model be reliable without thinking?
I like to think of it much like (as a common example most people can relate to) the Newton-Rhapson method for finding roots of a (mathematic) function. Your initial prompt runs, then the ‘harness’ kicks in using whatever methodologies are behind them to iterate on that prompt (back and forth with the model, occasionally with the user to get better guidance) and refine the outputs to hopefully converge back to some sensible output or actions the user was initially looking for.
So you’re hoping for an LLM that sort of ‘zero shots’ or needs minimal iterations from a prompt to give usable results. I find from my anecdata it varies across models and what I’m trying to get it to converge on. I tend to prefer models to not zero shot attempt because they tend to not do great, I want them to get feedback often to let me push them down the route of convergence in spaces I already understand well, meanwhile I like them to explore and give me new paths in spaces I’m not too familiar with.
That’s really what all that “second guessing” is, it’s making sure you’re following a sane path in a massive parameter space of an ambiguously defined problem. Imagine if in Newton’s method you checked the slope and it didn’t decrease from the last iteration and you just say “screw it let’s keep trying that direction.” LLMs and their harnesses tend not to have that base assumption like iteration on decreasing slopes to guide them closer to convergence, it’s a lot messier.
https://huggingface.co/bottlecapai/ThinkingCap-Qwen3.6-27B
I didn't realise there are people out there unironically using crush
Works better than opencode (pi based) or anything else I have tried for my needs, and by far the prettiest and easiest to reason about what is going on.
But I will bite. What does pi do today better than crush for your use cases?
What you describe is a engineering harness problem.
If you, and i mean the royal you, actually read tge thinking traces you can see and figure out where its stuck
This means an effective harness would observe when the model is overthinking and step in with reasonable redirection, like increasing logging.
Llamacpp can set reasoning budget and message per reauest, so it can be dynamic.
Your complaint is "skill issue" based and will be resolved by people who do something ither than vibe code react demos.
I'm using kimi-k3 on a real high level dev & analysis task at the moment (with Gemini flash for fast implementation) and it is fantastic.
I think we have passed the point where frontier intelligence is commoditized.
Fable-1m-max-thinking output, despite being incredibly slow and expensive, feels like it bucked a trend towards superficial loquaciousness in their models that had been building since 4.6.
I'm a bit of a luddite when it comes to upgrading models, fable was the first one to make me give up Opus 4.6-1M-max.
If you look at the hiring marketplace, being just marginally better than your peers can be very lucrative.
If you’re competing on speed or capability as a company (or as an employee), you’re probably going to be willing to pay for the frontier.
Good enough with a lower price will win the masses
I care a lot about speed and efficiency right now. For my setup I would like to have 2-3 different model families. I've settled on GLM-5.3 (formerly Deepseek v4 pro 0813) for architecting, Deepseek V4 Pro 0813 for developing, and Gemini flash lite (any recent cheap model) for repo scouting. I'll add another one in the mix for reviewing (in this case Gemini 3.7) and that's all I need.
I've tried most models except Grok.
Qwen is too expensive IMO (Alibaba Cloud subscriptions are hard to come by and I'm not spending 50 euros a month for a tool, so 18 euros it is). If it ever becomes efficient enough to run locally I will definitely look back.
Claude is slow and expensive (the cache hit prices are absurd).
OAI is pretty good, I might add it to my arsenal seeing how cheap it is.
These opinions change every day. Last week I would've never picked Deepseek until I read about the pricing. even post aug 16 it's worth it (although it's getting close to gemini pricing).
Right now my costs are 12 euros a month (z.ai) + whatever deepseek consumes. This typically isn't more than 8 euros a week. 44 euros a month and I have a setup that is doing pretty well.
Anthropic's recent releases are wordy to the point of exhaustion. Every time I use opus recently I find myself wanting to yell "GET TO THE POINT" at a terminal, which is exacerbated by it being slow.
I just wouldn't bias myself that way, most people haven't really used local models. This stuff is pretty much all subjective evaluation, there's plenty of reasons for people to favor certain models or disfavor others.
Dude, GLM-5.3 released _today_.
The phrasing "I've settled on" is incorrect for this context.
I think some of these are excited, eager users always ready to hype up the new thing. The same crowd that previously would constantly push for a rewrite from angular->react->svelt->god knows what. Instead now it is on a 6 week cycle and about models/harnesses.
I think most of it is bot driven spam by the various labs. It's hard not to notice 3 month old accounts with very strong opinions about various frontier labs and little else.
Ultimately, I think some of it is legitimate shifts in who's in lead and what is the best. You gotta dig through a lot of crap to get to that, and I don't really know how to to.
Ultimately I'm saying is that I always applied a fair amount of skepticism about what I see in comment sections but these day it is extreme amounts.
I've assumed the same as well.
I also assume that many of the companies developing these models engage in benchmaxxing.
At my company we've developed our own internal benchmarks for evaluating LLM models as they become available. The benchmarks are tailored to our particular use cases but the utility and knowledge our benchmarks assess is still fairly universally applicable. I see wide differences between what our internal benchmarks report and what the major benchmarks do.
There was a whole lot of fanfare about how amazing GLM 5.2 was when it was released, but it was pure rubbish on our internal benchmark -- far behind OpenAI, Anthropic, Gemini, DeepSeek, etc. I don't know how to reconcile the fact that GLM 5.2 performed very well on some of the major public benchmarks, but consistently performs so poorly on ours. OpenAI models tend to dominate our internal benchmarks.
Then Terminal-Bench 3.0 came out (where the questions are new), and GPT 5.6 Sol received 34.6%, while GLM 5.2 dropped to a whopping 4.6%.
That's odd, since Fable seems to be the leader for the industry. Not cost-effective, but if Anthropic models get dominated by OpenAI in your internal benchmarks, this calls their validity into question. Separately, see the jagged frontier effect. [1]
I've been using GLM 5.2 at my day job (mostly Rust backend work ATM). Nothing that blows away the models from OpenAI and Anthropic, but solidly good enough to get it done. A lot of people have experienced this and the fact that an open weights model can do so is where most of the excitement comes from. Optimizing for benchmarks can only get you so far, and people are quick to criticize models that fall into it (like DeepSeek Pro V4 recently).
[1] https://mitsloan.mit.edu/ideas-made-to-matter/working-defini...
Really no different than when there was suddenly online personas everywhere hyping up TSLA out of the blue. You can see the same thing going on with the BoringCompany subreddit. Crazy that the botnet master isn't able to convince us that Grok is also the best model. I don't think buying twitter was an accident it was probably just literally covering up the evidence.
https://www.rhsmith.umd.edu/research/twitter-bots-boost-tesl...
One thing I have noted a lot more of is that there is comparative fanboying going on. Like "this model has done badly, my favoured competition has a model out soon that will beat this in every way" — comparing a released product to unverifiable hopey claims about an unreleased product.
It's tempting to assume that is bot stuff, but if you've been around any other "hot" technical hobby online (cameras, phones, 3d printers, whatever) you will know it's not. It's just fans aligning into teams, some of them laconic and amusing, some of them overkeen and toxic.
Which itself released yesterday? You're writing, reading, and evaluating enough software in a ~36 hour period to form, reject, and form another opinion about which model makes better _architectural_ choices?
What you're currently doing is "testing out"
Likewise I have a tool-use eval set and a browser-use eval set. I use frontier models for most interactive tasks that are not home assistant or background agents.
I can imagine someone could build evals for that but I have never done so.
no they aren't. they discontinued their always-sold-out coding plan and launched QwenCloud (basically a friendly frontend with Alibaba Cloud as the hidden backend) and launched typical subscription plans for Qwen & co alongside it.
All of this DeepSeek price increase stuff is related to their specific API, not the entire market.
They were stealing your code for that extremely cheap cached token price.
That was never the real price. Every Chinese LLM API provider is the same to the best of my knowledge.
"CHEAP API" (but pssst we train on every single input aka your code and ideas are ours forever)
If some models are cheaper it's simply because they cost far less to develop. And/or run.
Most of the time Luna figures it out where Deepseek was failing. I rarely have to go to Sol levels.
And yes, OpenAI and Anthropic are highly questionable, too, and so I've ended my OpenAI subscription and I'm moving from Claude to other providers. But even people who do not do that are absolutely justified in drawing a line at supporting Musk, because there's a difference between mere normal shittiess and Musk.
So you get worse outcomes while also financially supporting one of the most prolific pushers of race wars, the destruction of labor rights, environmental protections, and general misinformation.
Truly a win-win?
Contrary to your framing it seems to only reason to use Grok at this point would be ideological.
No one actually cares about the politics as long as the model codes well.
Edit, quite interesting to see the reception to this comment compared to essentially the same type of comment I made on a Grok 4.6 benchmark HN post: https://news.ycombinator.com/item?id=49275385#49275571
It's true that Cursor gives a lot of usage with Grok, most users of Cursor don't care about Musk.
And yeah, people that don't care DO make the world worse through their apathy.
It is not all just 'politics'. Take a stand on some issues. It doesn't cost much not to use Grok.
Some people might not mind (or even know), but I sleep better at night trying to work as ethically as I can.
> I totally care about politics, especially when it comes to not giving my money to people like musk.
(Btw, mussolini didn’t make the trains run on time)
I do. There are enough models to choose from that I don't need to use one from a guy who did a nazi salute at the presidential podium, and proceeded to rid our government of brown people he felt didn't deserve their job.
I also miss the Qwen 3 Coder Next, which was 80B A3B, there are quite a few use cases where a non-dense model <100B would be the sweet spot (when you have the VRAM but not the TDP or compute power). Heck, I'd gladly take A5B or A8B or even A10B as a sort of middle ground.
Also alternate link for viewing the images without signing in: https://xcancel.com/Alibaba_Qwen/status/2088280182356611304
Qwen3.5:122b:a10b can run about twice as fast as this 27b dense model.
Edit: Like its predecessors, 3.8 seems really inclined to overthinking, and on a 27b dense model that's kind of painful. I think I'm going to stick with gemma4:26b-a3b as my go-to because it runs about 4x as fast and tends to only need a fraction of the tokens in its 'thinking' stage to get the same or similar answer.
https://huggingface.co/froggeric/Qwen-Fixed-Chat-Templates
[1] The specific directions are "Reasoning effort is set to xhigh. Please think carefully through the task, validate key assumptions, consider plausible alternatives, and prioritize correctness, consistency, and clarity in the final answer." and "Reasoning effort is set to low. Keep your thinking brief and focused, moving directly to the conclusion without unnecessary elaboration."
The MoE models are stupendously faster.
For comparison's sake, Qwen 3.6 35B A3B MLX, same prompt, same machine, thought for 1.59 seconds, used 2,398 tokens, and hit 80.83 tok/s. It actually output just a straight SVG file of an owl (not a bad owl either!), whereas Qwen 3.8 27B spent a zillion times longer gold-plating everything and output HTML+JS+SVG with interactive animation, sound, etc., etc.
Edit: link - https://huggingface.co/rico03/Qwen3.6-27B-Claude-Opus-Reason...
https://huggingface.co/Qwen/Qwen-AgentWorld-35B-A3B
KAT Coder is another Qwen 3.6 MoE fine-tune that also improves on 3.6 by a measurable amount.
But, I do hope for a bigger MoE, in the 70B to 120B range, something in the Coder Next lineage. I've got a Strix Halo that isn't getting used to its best ability because the best models all run fine on my faster desktop dual 32GB GPU setup, and they run too slow for comfort on the Strix Halo.
Edit: I guess you're right; apparently it's for simulation. I didn't look into it beyond the benchmarks. But, it does work in an agentic context, regardless. It'll write code, and drive an agent.
MoE is literally 5x times faster (on CPU) than comparable dense Qwen!
Tested:
Qwen3.6-35B-A3B-UD-Q4_K_XL
Qwen3.8-27B-UD-Q4_K_XL
27 dense is far more capable than 35A3.
llama serve -hf unsloth/Qwen3.8-27B-GGUF:UD-Q4_K_XL
but it failed my basic prompt to compose a vim regex to match CamelCaseWords
downgraded a bit with Q4_K_M from ollama run qwen3.8:27b
and /set nothink and at least 1 regex matched FooBar
prompt eval at 2.8 t/s eval at 0.94t/s
I particularly enjoyed this usage of the regex: /%\1\%/ ... onward for 700+ characters of \%\/ :-)
I'm using a Q4 quantization from unsloth (Qwen3.6-35B-A3B-UD-Q4_K_XL). It gets up to ~20 tokens/second in generation. I don't know precisely how much KV cache I can safely use, but it's in between 140k–256k. (I.e., 140k reliably works, 256k kernel-crashes from OOM. Don't feel like bisecting).
Inference is llama.cpp with the Vulkan GPU backend on Linux. (I.e., -DGGML_VULKAN=1 on the llama.cpp build, and --gpu-layers all on llama-cli or llama-server. (And for my specific setup, two kernel parameters specific to amdgpu: ttm.pages_limit and ttm.page_pool_size. A driver VRAM limiter. Look it up if you're on amdgpu!)).
I assume it would not be able to handle an unquantized Qwen3.6-35B or is it irrelevant as you almost always would want to run a quantized version of the model on consumer hardware?
on the 395+ AI MAX w/128GB, the A10B qwen 3.5 can do a lot of long running work if you don't need to baby sit it. deer-flow works well like that.
Whats up with focusing on the active param count? Do yall fiddle with the weights or something?
Dedicated GPU memory tends to be much faster (either DDR6 or HBM). So if you can fit a whole dense model in it, you're probably better off with that.
I saw unsloth has 1bit quants too so I might check that out, anybody have experience with those?
I’m working on a setup that’s more geared towards running tasks overnight so the slow tok/s doesn’t matter as much
And this is also specifically for the 1bit quant version. I don’t think the fp8 or even fp4 versions have this issue, but I haven’t tried those much
I've got an RX 7900 XT (20GB of VRAM) and I can run glimmer with a full 128k context window with the draft model at 65-80 tok/s.
This model, on the other hand, I get about 30 tok/s with a 30k context. Raising the context or loading the draft layers for MTP drops performance to 9-15 tok/s.
So I wonder how big the "real world" delta between Glimmer and Qwen is here. I can already run 3-bit DSv4-flash at 9-15 tok/s with 100k~ context, and I suspect it would outperform 4-bit Qwen 3.8 27B here.
I'll have to experiment and see if I just made a stupid mistake somewhere, but it looks like Glimmer might make more sense for the comically specific niche of "20GB VRAM".
that's using llama.cpp
llama-server \ -m ~/somePath/Qwen3.8-27B-UD-Q8_K_XL.gguf \ -np 1 --kv-unified \ -fa on --no-cache-idle-slots --reasoning-preserve \ --temp 0.2 \ --spec-type draft-mtp,ngram-mod --spec-draft-n-max 3 --spec-draft-n-min 1 \ --cache-type-k f16 --cache-type-v f16 \ --chat-template-kwargs '{"preserve_thinking": true}' \
I tried playing with all the recommended parameters from the unsloth page with no luck...
in one of the high fever ramblings it ended with amen... lol
I still need to understand that.
Setting a reasoning limit does not seem to have good results, because it really seems to go down rabbit holes and that means that cutting reasoning off too early is going to punish the quality on anything it has not got round to pondering yet. But maybe I have to give it a bit more room.
I have not tested in an agentic sense yet, just with my sort of pet queries in LM Studio, but it rather looks like it expects an agentic flow, because telling it that it's a helpful coding agent and changing the order of things in my prompts (telling it up front to ask any clarifying questions before detailing the rest of the prompt) has definitely kept its thinking a bit more on track.
> --temp 0.2
Looping is a common symptom of changing the sampler settings from what it was RL trained with.
llama-server -m Qwen3.8-27B-IQ4_NL.gguf --mmproj mmproj-BF16.gguf -c 170000 --parallel 1 -ngl -1 --cache-type-k q8_0 --cache-type-v q8_0 -b 1024 -ub 512 --flash-attn on --no-context-shift --no-mmproj-offload --spec-type draft-mtp --spec-draft-n-max 5 --spec-default --cache-type-k-draft q4_0 --cache-type-v-draft q4_0 --threads 24 --jinja --reasoning on -fit off
Identical to the qwen3.6 config. With a prompt like "svg owl" (which can reuse quite a lot compared with creative writing or similar, so ngram-mod shines), I get about 70-80t/s like this, with a memory overclock of about 1.5GHz
In my tests, even Q8 quantization for the KV cache comes with notable drops in performance for longer tasks. It does provide more context length in limited RAM budgets, but the longer context tasks are where KV quantization starts to show problems. It’s basically unnoticeable for simple and short tasks.
> --spec-draft-n-max 5
5 is a lot of tokens to draft. Are you really seeing acceptance rates to support that? When I tested it, 2-3 was the peak. Anything more started reducing performance except on highly predictable short outputs.
The thing is that I can either use the q8 context, or have not enough context window, so I just live with whatever degradation there is. The same can be said about the IQ4_NL. I would not go any lower though.
As for the draft count, indeed that depends on what you do with it, but for coding, reverse engineering and that kind of stuff it does pay off in my testing, though 5 is really pushing it, but the 4090 has so much compute.
Last logline I saw scroll by right now had 47% acceptance rate for 4th and 28% for 5th, but not sure how representative that is. I think when tuning 3.6, I saw more like 33%? But not 100% sure.
Though it might turn out that that doesn't offer any benefits. Has it been tried yet?
___
EDIT:
LLM pointed me towards this thing I'm not going to read
https://arxiv.org/abs/2512.11280
"AdaSD: Adaptive Speculative Decoding for Efficient Language Model Inference"
Some higher level nerd please go have a look at that. I have important meetings and golf classes to attend.
https://arxiv.org/abs/2607.05147
But they do so to maximise total throughput, I don't think there's reason to do that for batch=1.
The unweighted average is 4.1
So with qwen3.8 as well, I think I would not go below 3 in this setup with this workload. Meaning 4, and, to push it a bit further, 5.
Also, cool that it carries over from qwen3.6 like that.
Here is the chat template I used (and renamed to qwen.jinja): https://huggingface.co/froggeric/Qwen-Fixed-Chat-Templates/r...
For image support, drop to NGL 24 and add:
And drop the --no-mmprojArguably, what I am doing is also very very close to that, with the only difference being that I am somewhat less of an idiot than the average internet dweller you'd get on such a site. Or rather a different flavor of idiot.
Ideally, the people building the tools build them in a way that just does the right thing - which I am confident that llama.cpp does or will do in the future.
So you encode that knowledge not in language and online comments but in code and with a filter for actual expertise.
And, frankly, there's really not all that much to it. It's like maybe 3 parameters to play around with.
The valid solution space is pretty small, but people will want to make it "theirs" regardless, so you get non-solutions just so that everyone could also be a part of it. The usual social dynamics foo.
Let me return the favor:
Deepseek V4 Flash 0731 Q8
This is on a 4x3090 box with 256G RAM and a 5955WX TR; the CUDA drivers are patched to allow the 3090's to talk to each other. MB is a WRX80E-SAGE (which has 7 slots at x16), BAR > 4G on and PCIe BAR resize set to AUTO in the bios. OS is Ubuntu 25.04.
GGML_CUDA_P2P=1 build/bin/llama-server -m ../DeepSeek-V4-Flash-0731-UD-Q8_K_XL-00001-of-00005.gguf --alias DeepSeek-V4-Flash --threads 32 --threads-batch 64 --cache-type-k f16 --cache-type-v f16 --metrics --cache-ram 0 --load-mode none --fit on --fit-ctx 1000000 --host 0.0.0.0 -fa on --top-p 0.95 --temp 0.7 --repeat-penalty 1.2 --jinja --tensor-split 1,.8,.1,.15 --cont-batching -b 4096 -ub 4096 -ncmoe 36 --parallel 1 -ot 'blk\.3[67]\.ffn_(gate|up|down)_exps=CPU'
I'm still working on an optimal packing for the draft variation on this theme but that gives all kinds of issues. The 1,000,000 context is very nice to have.
Searching through Reddit and forums for best commands is annoying.
But maybe someone else has experience to share there
put differently, if you put a random person infront of whatever model you used (say, a 50yo receptionist at a pharmacy in india), they would not have been able to create that, because they would have lacked the motivation, idea, background knowledge, taste, etc to create such a thing.
I sure hope my boss doesn't think he built my work! He'd probably get fired pretty quickly during on call!
Your boss is a human.
This is a computer program running on your PC.
I hope you can see the difference.
Most managers do though?
But that's not the same as me being needed to be on a call with integration teams and having to run the call. It's just a layer of abstraction for management. I own the creation side of the product.
Who created it is still me.
None would claim they chiseled anything, if it was 3D printed. They may claim they designed something.
Legal questions of authorship are going to have to be established in terms of doctrines like SSO [0] and AFC [1]. Currently the incredibly sparse caselaw around this has yet to involve such non-literal notions of copyright.
[0] https://en.wikipedia.org/wiki/Structure,_sequence_and_organi...
[1] https://en.wikipedia.org/wiki/Abstraction–filtration–compari...
Absurd logic. The AI built the website.
Just like if somebody creates software today, and the end result is generated by the collective effort of thousands of agents, the "Director" still gets credit.
By which I mean it puts the llm in the role of actor and such… people practicing an art. It seems harder to make a case that the llm is actually making creative decisions rather than aping a synthesis of past human decisions and expressions of taste, strictly, at every turn.
Though I guess it’s all “shades of grey”. I never felt like I was contributing a ton when I was essentially wrapping a large, capable library that some smarter person or people made, either.
Tangentially… remember back when we used to see headlines like “12 year old made a web browser!” and we all thought, “No. No they didn’t.”
It all feels a little like that, again.
How many films do you remember the producer above the director?
If there were multiple people involved in the creation of a film I helped to create, I cannot factually say I created it. Just like if someone builds something using code generated by AI, they can't factually say they created it.
Lol, yeah I'd wager AI tooling skills are less relevant than knowing your way around a shell. It doesn't help that all of the AI tooling has event more churn than js libraries and package management systems!
I think there's probably some value in understanding how LLMs work, but beyond that it's the same boring skills that matter the most... Critical thinking, design chops, attention to detail, perseverance, etc.
For example, of his 1500 spot paintings, he only actually made 5 of them.
It's not uncommon at all for artists to work this way.
llms are not human.
$ build/bin/llama-server -m Qwen3.8-27B-IQ4_NL.gguf --mmproj mmproj-BF16.gguf -c 170000 --parallel 1 -ngl -1 --cache-type-k q8_0 --cache-type-v q8_0 -b 1024 -ub 512 --flash-attn on --no-context-shift --no-mmproj-offload --spec-type draft-mtp --spec-draft-n-max 5 --spec-default --cache-type-k-draft q4_0 --cache-type-v-draft q4_0 --threads 24 --jinja --reasoning on -fit off
0.02.993.689 E ggml_backend_cuda_buffer_type_alloc_buffer: allocating 911.53 MiB on device 0: cudaMalloc failed: out of memory
Update: Oh, it works after I stop Xorg. But nvidia-smi only showed Xorg using 200M out of the 24G, so why would a 911M alloc fail?
If you're trying to use the 4090 for your desktop at the same time, you will need smaller context size. You could remove the '-c 170000' and try turning fit on to see what fits.
That was just the last buffer allocation request that failed, it didn't tell you by how much it failed by. It could have failed it by a few kilobytes, it could have failed it by 910MB. One would guess it probably failed it by a couple hundred megabytes in the end judging by your results.
[1] https://github.com/ggml-org/llama.cpp/releases
Instructions if you want to do the same:
1. download two files llama-b10434-bin-win-cuda-13.3-x64.zip and cudart-llama-bin-win-cuda-13.3-x64.zip from that llama.cpp Github releases page, and extract both into the same folder.
2. Download the Qwen3.8-27B-UD-Q2_K_XL.gguf file from huggingface and put it into the same folder beside the `llama-server.exe`.
3. Create a file named "RUN_QWEN_3.8.bat" next to `llama-server.exe` and put the text above into that bat file. Double-click the bat file, then open http://localhost:8033 in your browser to see a chat window.
You can use it with any agents by pointing them at http://localhost:8033/v1 which is a working OpenAI compatible endpoint (it doesn't use a token, if you give one it's ignored).
Congratulations, you're now running Qwen 3.8 27B.
Note: I built the computer in question for playing games, yes it needed to be Windows 11 for anticheat reasons to play games with family, I didn't want to dual boot so here I am. I figure I should share instructions for folks who may also have a Windows PC around for such purposes. Specs for this are AMD 9800X3D, 32GB of system RAM, RTX 5070Ti 16GB
From: https://github.com/ggml-org/llama.cpp/blob/master/tools/serv...
If you've tried it, what worked well and what didn't? I'm especially interested in mismatched VRAM setups, e.g. a 16 GB GPU + a 24 GB GPU.
How much overhead did you see from inter-GPU transfers, and did the extra usable VRAM outweigh the performance hit?
Llama is decent at auto optimizing it if you let it use both gpus. It’ll split the workload so the contiguous layers are all on one gpu. Once the model is loaded, you only transfer weights between gpus once (per token?), at the layer boundary.
I run on an 8gb 3070 and 12 gb 3060, and the only weird thing is that the weaker card gets more layers (and therefore work) because it has more ram.
Oh, if you’re barely fitting the models into your vram, you may need to explicitly adjust the layer balance between cards — sometimes it fails to realize it should have put certain things (like draft models) on the other card so you can fit one more layer in.
To provide some anecdotal data, here is how my 5090 + 3060 setup performs with Qwen 3.8 27B (Unsloth's UD-Q4 with MTP):
For reference, here are also some numbers from my 4060ti + 3060 (16GB + 12GB) setup. [0][0]: https://news.ycombinator.com/item?id=48700091
I use an RTX 3090 (24GB) and a GTX 1080ti (11GB). Just the 3090 for 3.8 I get maybe 60tg/s (UD-Q4 quant), for both I get around 40tg/s (UD-Q6). Not apples to apples though considering it's different quants.
Here's my config: https://gitea.va.reichard.io/evan/nix/src/branch/master/modu...
Disregard references to RTX 6000 cards, most of it is generally applicable to all multiple-GPU boxes.
I just want to run `<some-command> <model-name>` with some default parameters set and for it to run locally.
It's long, I guess, but not cryptic.
You tell llama server where the model is, which context size to use, what to use for the K/V cache quant, that it should do MTP, tune some MTP parameters, and that's kinda it.
Perfectly logical blocks with all the model-specific weirdness (that does exist!) abstracted away.
You could also just run -m <modelfile> and let llama-server do the right-ish thing. The defaults are probably fine, but not how you squeeze out these exact numbers. I think at least. I've never tried. My hubris stopped me from trying auto configs.
isn't that the hard part? You know the ballpark ideal values for these many parameters since you're a knowledgeable expert but the vast majority of people are just like "I want AI" and have no idea what all the jargon even means.
Otherwise, if you're a programmer setting up a local harness, it only takes like 20-30 minutes to learn what the right parameters are.
It's very model, hardware, and use case dependent which is why a one size fits all solution doesn't work
They could ask your current agent to a) search for this type of content online for the optimal setup for their hardware b) have the current agent/harness spin it up have it verify the config run few experiments.
Sure AI may make mistakes, or won't get the best possible config probably, but it certainly do a good enough setup, this is a task with feedback on whether the server crashed or poor performance easily measured so the agent can do a pretty good job.
Because this is kinda the one new thing that arrived in the technology scene, so getting at least some amount of understanding of its "inner" workings might prove useful in the future.
Beside that, it is also just.. interesting? It's fun tuning the machine to see it improve. For some, anyway.
The pain point they raised is this is too complicated for people who just want to get started, that is not true anymore.
It is certainly fun to fine-tune and setup if you like do something like that, however the need to do it hardly is a barrier for those who don't want complexity as OP imagines.
Lower level API/interfaces should not be a barrier for people if they are apply framing that way. More and more people are thinking agent native so this is not really a issue.
Why this headache inducing lingo tho? What does that even mean, and why should I sign up for your webinar about that?
For Claude, I setting a single config file and then download and run Claude Code CLI. Even easier for the GitHub Copilot CLI.
There are some clients that will index the models and allow you to do that but I'm no expert, I've used OLama studio but it always seems to go weird for me.
Even this command above, it's not clear where op got the model from. So I'm with yah.
For example, op uses : Qwen3.8-27B-IQ4_NL.gguf.. But I cant see where to download it. It's not tagged on hugging face at least..
If I want to download the model myself, it's not clear. I thought it was supposed to behave like a package manager. But even in nuGet I can download a zip of the package.
They shared a lot of links, I'm struggling to find yours. Where did yours come from ?
Who is Unsloth AI? Have they modified the model ? Is this really the source of truth ?
Do you see how steep the barrier for entry is to do anything right ?
unslothai is not a name qwen has ever used. So you're sharing a link to a model that isn't from the owner, while saying it's the owner's. I'm not comfortable with that, and I want AI to be a better tool.
You asked where to find the GGUF files of this model for direct download and I provided it. Almost all useful model files that can be downloaded are hosted on Huggingface.
> They shared a lot of links, I'm struggling to find yours. Where did yours come from ?
I went to Huggingface, went to the Unsloth org, as they tend to be the best, went to the Model page, and went to the "Files and versions" tab.
> Who is Unsloth AI? Have they modified the model ? Is this really the source of truth ?
Unsloth AI is a very popular, highly reputable organization that takes upstream model files, performs some optimization, and provides models in various formats. Apart from speed tweaks, they do not modify the models. They also provide useful benchmarks, copious documentation for local execution, and a Studio application for easy execution and post-training of models.
> Do you see how steep the barrier for entry is to do anything right ?
No. Searching for this information is not difficult. The llama.cpp documentation and guides that Unsloth provide are all you need. Search engines can take you further if you want.
> unslothai is not a name qwen has ever used. So you're sharing a link to a model that isn't from the owner, while saying it's the owner's. I'm not comfortable with that
Qwen also provides models in GGUF format on Huggingface, but they will not be as performant. Even when first-party GGUFs are available, most people will prefer quants from Unsloth or a few other popular optimizer accounts.
> I want AI to be a better tool.
Best of luck. Your attitude and unwillingness to even try and learn on your own when people have tried helping have burned my good will, and this is as far as I'm willing to carry you.
"I went to youtube, gmail, ycombinator, deliveroo, then I went to another site I randomly chose, because they're the best, duh"
Okay.
> Unsloth AI is a very popular, highly reputable organization that takes upstream model files
And How am I supposed to know that arriving to huggingface as a new user? Enlighten me.
> Qwen also provides models in GGUF format on Huggingface
Cool.. Why don't they share em because I genuinely cant find em, I'm dumb.
> Your attitude and unwillingness to even try and learn on your own when people have tried helping have burned my good will
Well, I'm willing, but not from people who I might burn good will. Gracious. You do you.
The path I described is all within HuggingFace.
> And How am I supposed to know that arriving to huggingface as a new user? Enlighten me.
Because I told you, knowing it was the best starting point for newbies.
> Cool.. Why don't they share em because I genuinely cant find em, I'm dumb.
You could go to Qwen's organization page on HuggingFace, it has a search function at the top, but you would be better served sticking with Unsloth.
> Well, I'm willing, but not from people who I might burn good will. Gracious. You do you.
Expecting others to do everything for you is not the same as trying things and asking questions about what you found.
It's just not. Where do I see the model download ?
The fact we're here is a loss.
It's not intuitive. Deal with it, or fix it.
Huggingface, Unsloth, and llama.cpp all have documentation you can follow that will exceed anything I can tell you here. LMstudio, Lemonade, or Ollama might be even easier for you to use. Take my suggestions or don't.
You don’t need to fine tune all of those parameters to get started.
It’s really easy to ask an LLM to adjust the command line if you can’t be bothered to read the help out. Copy the help output into the LLM and tell it your goal.
> Ollama is confusing and doesn't seem to support Qwen3?
Typing “Ollama qwen3” into Google takes you right to this page:
https://ollama.com/library/qwen3
If even Googling for basic Ollama support is too hard, there might come a point where you have to acknowledge that local LLMs are not for you. None of this is really that hard with some basic Google bootstrap skills or by asking an LLM to help with the command.
> Attention: To be updated for Qwen3
on Qwen's official docs: https://qwen.readthedocs.io/en/latest/run_locally/ollama.htm.... It's not like I just made it up. Of course I searched "ollama qwen3" and saw what you linked, but that doesn't mean it "works". I have other things to do besides to try a bunch of poorly documented and executed tools just to see if it works or not.
I guess the TLDR is that I'm stupid or lazy. Also, everyone is responding about how easy it is, and yet, it's apparently so easy that it's hard to document well.
You could override, obviously.
Woa, is that still a thing? You mean like SOCKS5 stuff that you have to manually configure in every application that uses the internet?
I mean maybe I'm just living under a rock but I feel like that's a rather niche situation you got there.
Every big company in the world uses a network proxy. LM Studio, as far as I can tell, cannot be configured to work behind such proxies.
It's becoming more rare, now.
A lot of the universal truths about corporate networks from the early 2000s are no longer true today. Some companies are stuck in their ways though.
The overlap between companies that require someone to use a network proxy and companies that have GPU-equipped machines with enough RAM for LLMs and and that allow people to download and run executables of their choosing has to be small.
I would love to see more data on that because I've seen it constantly. There is more isolation maybe where you can do whatever on 'open' network, but always some kind of proxy/vpn connection for hitting anything sensitive.
The operlap is there.. But I would be worried if it was just flat out taken away from secure managed connections just because of AI.. Again, would love to see the numbers of your assumptions.
Makes it rather weird that LM Studio doesn't support it given how their target market, or well at least for their paid products, is very enterprisey.
I remember there was a short story in BYTE Magazine about a similar kind of scenario way back when, I think at least 30 years ago, long before LLMs and AI agents became a reality.
I am benchmarking it now locally, will put the results and speed/tps on aibenchy.com
There's some YouTube guides for it.
I also undervolted my new 5070ti, same tdp, around 260w instead of 300w and like 8% better performance.
Chinese (small model) AI is 6 months behind American (largest model) AI.
Unsloth's GGUF quants are up: https://huggingface.co/unsloth/Qwen3.8-27B-GGUF
I'm a huge open model fan, and have used them since forever, even have daily drivers for on-prem dev, but no. They do not beat opus on real-world usage.
Qwen models are impressively good for what they are, are "good enough" for plenty tasks, can be ran locally on decently priced hardware, and so on. They certainly have their uses, and the field in general has advanced faster than my early expectations. But to compare a 27B model to SotA behemoths from a few months ago is doing everyone a disservice, especially people who pick it up, try to use them just like API models, and leave disappointed and confused. Number goes up on a benchmark isn't it.
We have an internal eval that measures performance on tasks for a handful of embedded systems repos for our mmWave radios (mostly Rust, some C for microcontroller stuff). Qwen3.6-27B scores only 4% lower for pass@1, n=250 compared to Opus-4.8.
For the labeled dataset, the average PR size they're being measured against is around 1.5k SLOC.
This is very much "real-world usage" for us. The sort of change sets that come in daily/weekly and are solving non-trivial issues in the respective codebases.
As is usually the case, the most broad claims from both the labs and from the consequent pushback are talking past each other.
0% is 4 percentage points (pp) less than 4%.
I agree, but then we just need meaningful benchmarks that clearly show that! Otherwise it's hand waving about something that should be put on paper in quantifiable terms.
That's the rub. AI benchmarks are IMO, by and large totally unreliable. We think of them as similar to traditional benchmarks of deterministic processes where the number of variables is low. But they're anything but that. Non-deterministic processes with an astounding number of variables and fuzzy acceptance criteria.
It leads to results like these, where if you take it at face value, the only conclusion you can draw is "wow Anthropic must be stupid if Opus takes 1T parameters to do what Qwen can do in 27B."
There are tools like promptfoo designed for this.
Ideal local model would not know stuff like who Britney Spears is, best to leave precious weights for something useful.
Of course the line is very blurry but I'd be perfectly happy with local model that doesn't know anything about history, geopolitics, art or even biology etc. just coding, operating systems etc.
For example you would tell a model hey, become an expert in this language for me, search it online, it would still need to learn it and download the data to it's context and then increasing the memory usage, there's no way around it.
deepseek-v4-flash needs web search to return true facts.
I can say I find the law stupid, so no one should say person A beat person B in court. But I did not prove the law is stupid; I merely thought it subjectively and demanded others to follow the second part because I believe the first part is true.
Saying that "if the law is useless, court cases are useless" is objectively true and cannot be argued with. But you still need to prove why the law is useless, not why you think it is and even then if people disagree and use the law as a reference, then it's not objectively useless and court cases are not useless as well.
I am just now getting the benchmarks running against 3.8 27b but I expect similar results from benching 3.6 27b at the same quant.
So advantage is not having to produce your own quantisation / gguf from .safetensors you've linked.
Run the one you linked if you are running vllm (safetensors)
It's a bit bare at the moment, I assume they are going to add further detail later (eg comparison to other quants), similar to their other releases.
Sometimes they're just slow and expensive, so we we KLD as a proxy measure and it's very high correlation (95%+)
Gemma 4 31B: "Um, if I really said all of that, I guess I'd say this next"
Gemma 4 26B: "Dude, I would've said completely different stuff" (large divergence)
Gemma 4 12B: "Umm, there's zero chance I would've said some of this" (INFINITE divergence)
Gemma 4 E4B and E2B: "Derp derp, I'm happy to say almost anything" (lowest divergence)
For models which are chat trained, they simply would not recite Wikipedia, so the divergence is almost meaningless. I thought about capturing a realistic coding session and trying to use that as the corpus, but you need to preserve the turn-based tokens and such, so I moved on to other things.
Just wanted to say that this is a very important point that I totally agree with. People are obsessed with KL divergence, but it is yet to be demonstrated to be a descent proxy for agentic coding benchmarks.
It also has a lot of resolution and not a lot of noise. Better would be multi-turn benchmarks with tools but getting good precision and accuracy for that is hard and computationally expensive.
I am currently using Qwen 3.6 on RTX 3090 and I have to admit that without MTP it would be too slow to be acceptable for me (30-35 tok/sec without MTP, 60-70 with MTP). Without MTP I would just use OpenRouter and rather pay for speed despite having a capable local setup.
e.g. If you try to chat to it about something philosophical for example, or maybe a debate / creative writing, then you'll very quickly see how it is still a much smaller model at the end of the day.
Still, it's such a relatively accessible model to run, and I find a big part of leveraging smaller models is to give it well-scoped tasks; not too high level or ambitious ones. Very impressive for its size and the ability to run locally :)
I'm benchmarking the 3.8 model now, it seems it is better (near-perfect score on my E-mail spam filtering benchmark, best of any model I tested, ever). But it is slow.
One thing I would recommend is keeping an eye on MTP parameters. I tested and benchmarked extensively, and I use `--spec-draft-n-max 2` with llama.cpp. Longer sequences actually decrease overall performance.
As for running, I ended up using llama.cpp and its llama-server, with a bunch of scripts written by AI, because I got tired of LM Studio not implementing the image-related parameters which made gemma4 useless for OCR.
I am looking to automatically recover mail from junk folder due to bad icloud mail filtering.
Unsloth: https://huggingface.co/unsloth/Qwen3.8-27B-GGUF
This might work for you, but I didn't get on very well with MTPLX when I tried it a while back; YMMV: https://huggingface.co/Youssofal/Qwen3.8-27B-MTPLX-Optimized...
2 bpw - ~85.7t/s
3 bpw - ~58.0t/s
4 bpw - ~43.9t/s
6 bpw - ~29.5t/s
8 bpw - ~22.2t/s
16 bpw - ~11.2t/s
without cheating. You'd have to exclude layers, skip operations, etc. basically do stuff the model wasn't trained for. And speed collapses so fast with context that even 2 bpw would be looking at ~37.6t/s after just 128K tokens.
MTP only improves the situation by up to 2x in the ideal case, while drastically reducing the performance floor. While optimizing a 9B model on this hardware, I've found that the GPU just doesn't have enough FLOPS to handle speculating more than one or two tokens ahead on a single stream, regardless of quant level, simply because of the arithmetic cost of the forward pass. The 27B model would be even more expensive than that, potentially such that it's already bottlenecked by the GPU itself rather than memory.
I wouldn't get my hopes up for the 35B-A3B either. Not only is it reportedly much less intelligent, but I hit a similar ~85t/s wall in practice (again with highly specialized inference).
Without speculation I can reach around 120t/s on Qwen3.5-9B and with n-gram speculation (not even MTP; this derivative didn't come with one) around about 150t/s on average. This is on the very very edge of what I'd consider acceptable for me to even consider using such a compact model. YMMV due to the silicon lottery but the situation isn't good.
Also whats your cutoff for 'acceptable' speed? I would have said 25tok/s.
I consider 'acceptable speed' to be around 150t/s. Why? Well, this is generally what it takes to keep me engaged with the output, rather than immediately switching to other tasks and checking back later. When I check back later, I have more catch-up to do at once, and I haven't been following the process. So I have to recall it, familiarize myself with the new progress, and sort of get back into focus with it, which is a lot of mental work (even if it happens quickly in real-time). I prefer not to have to do this because of how much work it is, so I prefer to watch the agent in real-time and try to follow its reasoning. That also lets me interrupt it quickly when I see it about to make a mistake, or see an important detail I left out.
Most of the benchmark improvements afaict are in agentic and instruction following benchmarks.
To the countrary, the reason Chinese models are excelling in the smaller area is because there's tons of fat in closed source models because of the crazy cash being thrown around.
There absolutely is space to improve intelligence and capabilities without lathering on more and more parameters.
For FP8, on a single Nvidia 6000 RTX 48GB, i get 13 tokens/sec on a single GPU and 46 tokens/sec batched.
And I know the 96G nVidia cards are selling for over 10k$.
The future can’t arrive fast enough!
If you want Qwen3.8-27B Serving Configs for the DGX Spark vLLM NVFP4 and RTX 4090 llama.cpp GGUF I added the setups here https://x.com/ErdalToprak/status/2088299678085308761?s=20
Insane if that is the case. Downloading now!
qwen3.8:27b-mlx
Opus at home
I hope there also will be a new ~10b variant
Small local models can also be used for sub-agent tasks in most agent harnesses. But I'd probably run a larger MoE for that; they're faster and have broader knowledge. The dense models, even very small ones, are not blazing fast.
I don't code with any models small enough to run locally, at least not so far. Qwen 3.8 27B might be the tipping point, though. It's looking really promising, though it's probably slow enough that I won't ever actually use it. I'd rather pay $100/month for a faster model, even if Qwen 3.8 turns out to be smart enough for most of my work. Running it locally with the 8-bit quantization is going at 12-30 t/s, depending on how much context it's chewing on. So, if all you do with AI is coding, then you're better off doing it in the cloud.
But, there's lots of things a small model can do that aren't coding.
categorization, information retrieval, semantic search, image description
also with the model as part of an agentic system with tool calling
(edit: it is quite impressive what a small model in a feedback loop can do)
At rely health, we use OSS models in a HIPAA complaint and SOC 2 complaint environment to take advantage of asymptotically $0 intelligence to provide best in class care navigation . This should be industry standard -
It's still very useful, and it'll probably displace a good bit of API spend; but it's not really "trading blows with SOTA from just a half year ago". A bit overblown on the Anthropic/OpenAI has missed the window I reckon.
Also, on the open weight frontier side, Kimi K3 is pretty expensive, and Deepseek V4 Pro/Flash is getting a little less juicy with price increases.
Intent understanding is a big point for sure, but world knowledge I'm not sure I see a use case for it.
This is without saying that company clients of Anthropic or OpenAI will not have the upfront money to get a RTX 3090 for each and every employee. Much better to sign a deal with the LLM providers and get cheaper token price with a sales person.
That’s quite inefficient. Get 2x RTX 6000 Pro and run DeepSeek v4 Flash for the whole company. This may or may not be cost effective compared to cheap third party providers, but it lets you own your own data, is not actually that expensive, and reports suggest that you can get 2k tok/s or so from a setup like this (at respectable batch size, but that’s exactly what you want in this situation).
I’m not sure I believe that the unquantized performance is as good as people are claiming, but even light quantization would make it fit easily.
Nevertheless, I doubt anyone's going to emerge unscathed when the valuations start falling though. Vast sections of the American stock market will be wiped out yet and with it the global economy.
There will surely be tons of new companies that make a business off of hosting these models and even tuning them for specific purposes. Not to mention the massive pricing difference will benefit people who actually utilize them
Secondly, even with that being true, I think open models are a great net benefit for all mankind in the long run.
I think both of those things are true at the same time.
1. They still have extreme revenue growth (and despite what HN says, good gross margins supposedly).
2. People will pay a premium for the best model. So far, we see catchup but no leapfrogging.
Even Zuckerberg isn't as deluded to think that there's money in constantly releasing closed models and hypserscaling their deployment, which is why we got muse-glimmer 30b last week. I have been playing with it on a 32gb MBP w/ M2 Pro. If prosumer-grade hardware from almost three years ago can now offer a novice a way to host a possibly serviceable software development agent, then there's little incentive to pay out for LLMs like Anthropic's or OpenAI's over the long run.
Perf improvements seem to all come from training?
So one open weight model might "meet" Opus or whatever on benchmarks, but then fail to follow a simple answer format and also tool call correctly. The models are whipped to within an inch of their lives to strictly adhere to their post training quality gates.
I’ve heard rumours about AI shops optimizing for benchmarks. I also don’t think Qwen/Alibaba would be crazy enough to claim something unless there is some truth in it. Would love to see a side-by-side with Opus 4.6 on categories where Qwen 3.8 27B aces it.
Funny side note. It successfully one shot the program, but it wasn't able to run it because there literally wasn't enough VRAM left to allocate CUDA memory. Watching it try to debug that was fascinating. I'm pretty sure it would have killed the llama-server (and thus itself) if it hadn't been running in a separate container.
Completely local use is a different story, of course.
on my dual 3090s qwen 3.5 27b was running at around 110tps using the config from https://github.com/noonghunna/club-3090
make that 200tps on a single 5090, 4x faster than opus https://x.com/radixark/status/2088285681131110446
devs about to get handed a two 5090 box each and told to max that out
Couldn’t afford claude pro so I built an web based DSA coach that coaches you on DSA and System design in a socratic way. It uses a qwen 1.5B coder model and inferencing is all done on a CPU. ( who needs a GPU anyway )
https://openrouter.ai/qwen/qwen3.6-27b
Does it come down to utilization and/or specific model tricks and efficiencies (attention, kv cache, etc.)?
DeepInfra prices:
Qwen 3.6 27B: $0.32 in / $3.20 out
Gemma 3 27B: $0.08 in / $0.16 out
DeepSeek V4 Flash 0731: $0.08 in / $0.18 out
Qwen 3.6 35B A3B: $0.10 in / $0.95 out
https://openrouter.ai/qwen/qwen3.6-27b
https://openrouter.ai/google/gemma-3-27b-it
https://openrouter.ai/qwen/qwen3.6-35b-a3b
https://openrouter.ai/deepseek/deepseek-v4-flash-0731
dsv4f appears bigger and more memory/computationally expensive to run any way I look at it.
Just speculating, but look at attention. 35B-A3B context is heavier -- about 20GB per 1M tokens vs ~7GB. So if you have ≥10M tokens of context per inference node, then DS4 Flash uses less memory than Qwen 35B.
Qwen attention is probably more computationaly expensive as well.
Also I guess not much interest/competition in serving Qwen edge model.
Just tried using Pi Agent and looks very promising
https://avilpage.com/qwen-3.8-27b.html
I did notice if you go beyond Medium he starts overthinking like hell as per usual for a Qwen model.
Going to give it some coding tasks and see how it goes.
We been eating good at LocalLlama this week.
https://x.com/sgl_project/status/2088281320422322413
I only trust those users genuine personal tests
running an untouched, vanilla 4-bit version (Q4_0) I baked myself today (benched it against Q4_K_M (16gb) and IQ3_M (12gb), Q4_0 (15gb) is king)...
this model--
1: over 60% faster than qwen 3.6 version of the same dense 27b model, same engine setup (don't ask how, im not sure either)
2: has better reasoning quality, less "loopy" with its thinking patterns.. most certainly the smartest model on my roster currently
3: has the longest task horizon ive ever experienced (locally or otherwise)...I sent it a bunch of compressed ideas for an app, it sent me back the largest python app ive ever seen in one single ai response pass (80kb text file)
thanks qwen!! hoping to see the full model range get released...
https://www.youtube.com/@lukesdevlab
I don't know if that is what you are looking for or not and as always your experiences may be different.
However, watching tests of heavily quantized models that weren't designed for it (non-QAT) is frustrating. There's no way to tell if the actual model fails because it's dumb or if the lobotomy made it that way.
Was considering adding a LoRa/vision head to Flash, but seems like it could take a while to get it right.
If DSv4 Flash was multimodal, I’d probably be done model shopping for a while
Yes, I have a very dumb flow. The harness has a describe_image tool that takes an image and a prompt and so DSv4 Flash uses it to get an idea of what it's looking at.
my prediction was way too far out. 4.6 at home! Woo.
At least as compared to 3.6 27B in the same quantization.
Been running quantized 3.6 at 110t/s on a cheap 5060Ti and quite happy with it. If 3.8 improves on it, it would be awesome.
https://news.ycombinator.com/item?id=49299684
What a weird mechanism. If someone is judging a thread/topic/event impact by the number of points it got, then doing this unfairly degrades that thread.
It should have deduped by user and combined the 168(at the time of writing this comment) + 289 points. Just add the twitter link from the previous thread as an additional link in the description, like you normally do, move all the points over, and remove the old thread.
" The user wants me to explore the repository at XXXX and report back. Let me start by understanding the project structure, reading the CLAUDE.md file, and getting a general overview of what this repository is.
Let me start by reading the main project documentation and exploring the directory structure.
I'll take a look around this repo. Let me start by getting a lay of the land.
read resource CLAUDE.md (ctrl+o to expand)
ENOENT: no such file or directory, access 'XXXX/CLAUDE.md'"
This is the perfect candidate for just splattering it on your nvme and then reading it off there and into memory. All of these run perfectly fine on simple m4 silicone:
https://github.com/drumih/turbo-fieldfare
https://github.com/leonickson1/Swiftlet
https://github.com/sqliteai/warp
Even the fastest Apple Silicon chips like the M5 Max and the M3 Ultra still put up worse GPU compute performance than last-gen laptop RTX 4080 chips. And they don't scale, the largest M3 Ultra cluster you can configure is still ~2,000x smaller than a DGX SuperPOD. There's a reason Apple discontinued their rackmount hardware, there's very little demand for Apple Silicon in the datacenter.