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.<p>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<p>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
> --cache-type-k q8_0 --cache-type-v q8_0<p>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.<p>> --spec-draft-n-max 5<p>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.
Yes to both.<p>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.<p>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.<p>Last longline 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.
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 <a href="https://hypfer.github.io/will-it-fit-llama-cpp/" rel="nofollow">https://hypfer.github.io/will-it-fit-llama-cpp/</a> (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.<p>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.<p>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 generally agree and expect this to be the case from a legal perspective.<p>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.<p>[0] <a href="https://en.wikipedia.org/wiki/Structure,_sequence_and_organization" rel="nofollow">https://en.wikipedia.org/wiki/Structure,_sequence_and_organi...</a><p>[1] <a href="https://en.wikipedia.org/wiki/Abstraction–filtration–comparison_test" rel="nofollow">https://en.wikipedia.org/wiki/Abstraction–filtration–compari...</a>
You sound like your trying to reassure yourself of something.<p>I sure hope my boss doesn't think he built my work! He'd probably get fired pretty quickly during on call!
Do you find it useful or worthwhile to split a large LLM across two GPUs on a desktop?<p>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.<p>How much overhead did you see from inter-GPU transfers, and did the extra usable VRAM outweigh the performance hit?
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 <i>should</i> still be faster than the RAM?
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.
I just have one 4090 and nothing else, soooo I can't help there, but maybe someone else can.
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 <a href="https://github.com/local-inference-lab/rtx6kpro/blob/master/hardware/topology.md" rel="nofollow">https://github.com/local-inference-lab/rtx6kpro/blob/master/...</a> .<p>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.<p>I just want to run `<some-command> <model-name>` with some default parameters set and for it to run locally.
Start by copying the command line from the Unsloth guides.<p>You don’t need to fine tune all of those parameters to get started.<p>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.<p>> Ollama is confusing and doesn't seem to support Qwen3?<p>Typing “Ollama qwen3” into Google takes you right to this page:<p><a href="https://ollama.com/library/qwen3">https://ollama.com/library/qwen3</a><p>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.
What makes you say that it would be hard to do that?<p>It's long, I guess, but not cryptic.<p>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.<p>Perfectly logical blocks with all the model-specific weirdness (that does exist!) abstracted away.<p>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.<p>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 ?<p>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.<p>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.
Sure, but front ends like LMStudio exist for that crowd<p>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.<p>It's very model, hardware, and use case dependent which is why a one size fits all solution doesn't work
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.<p>For Claude, I setting a single config file and then download and run Claude Code CLI. Even easier for the GitHub Copilot CLI.
> LM Studio doesn't work behind proxies.<p>Woa, is that still a thing? You mean like SOCKS5 stuff that you have to manually configure in every application that uses the internet?<p>I mean maybe I'm just living under a rock but I feel like that's a rather niche situation you got there.
> I feel like that's a rather niche situation you got there<p>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.
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).<p>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.
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”?<p>You could override, obviously.
You know free LLMs can help you understand that command line or design your own...
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.
I hope really badly that we'll get a new 35B A3B or similar MoE model!<p>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.<p>Also alternate link for viewing the images without signing in: <a href="https://xcancel.com/Alibaba_Qwen/status/2088280182356611304" rel="nofollow">https://xcancel.com/Alibaba_Qwen/status/2088280182356611304</a>
I'm hoping too that they'll put out some MoE variants.<p>Qwen3.5:122b:a10b can run about twice as fast as this 27b dense model.<p>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.
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.<p>Edit: link - <a href="https://huggingface.co/rico03/Qwen3.6-27B-Claude-Opus-Reasoning-Distilled" rel="nofollow">https://huggingface.co/rico03/Qwen3.6-27B-Claude-Opus-Reason...</a>
Reduce or turn down thinking:<p><a href="https://huggingface.co/froggeric/Qwen-Fixed-Chat-Templates" rel="nofollow">https://huggingface.co/froggeric/Qwen-Fixed-Chat-Templates</a>
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.
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.
> I'd gladly take A5B or A8B or even A10B as a sort of middle ground.<p>Whats up with focusing on the active param count? Do yall fiddle with the weights or something?
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
You can run these on CPUs at a somewhat reasonable speed.
Beats Opus 4.7 Max (w/ Claude Code) on DeepSWE (42.2 vs 40). Looks like Qwen's 27B models continue to pack some punch.<p>Unsloth's GGUF quants are up: <a href="https://huggingface.co/unsloth/Qwen3.8-27B-GGUF" rel="nofollow">https://huggingface.co/unsloth/Qwen3.8-27B-GGUF</a>
> Beats Opus 4.7 Max<p>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.<p>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.
> They do not beat opus on real-world usage<p>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.<p>For the labeled dataset, the average PR size they're being measured against is around 1.5k SLOC.<p>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.<p>As is usually the case, the most broad claims from both the labs and from the consequent pushback are talking past each other.
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.
Let us know when you have Qwen vs Qwen comparison stats. As long as there's not a regression, that'd be awesome.
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.
> ...but no. They do not beat opus on real-world usage.<p>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.
A wise man once said, "not everything that counts can be counted, and not everything that can be counted counts".
In the end, the only benchmark that matters is your own.
Only useful benchmarks are those you (in particular) don't have access to.
> but then we just need meaningful benchmarks that clearly show that!<p>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.<p>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."
How can you say this when you haven't even tried it yet? Is it just hypothetical vibes?
There is 0 shot you can make that claim about this model you have not used or downloaded yet
"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.
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.
Is there any advantage to using the model from Unsloth compared with <a href="https://huggingface.co/Qwen/Qwen3.8-27B-FP8" rel="nofollow">https://huggingface.co/Qwen/Qwen3.8-27B-FP8</a> ?
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.
Depends on what software/hardware you'll run it. GGUFs from Unsloth can run on pretty much every single potato; full weights need beefy gpus
Unsloth one is gguf for llama.cpp (and some other on-device engines).<p>So advantage is not having to produce your own quantisation / gguf from .safetensors you've linked.
We also made NVFP4 ones if that helps! <a href="https://huggingface.co/unsloth/Qwen3.8-27B-NVFP4" rel="nofollow">https://huggingface.co/unsloth/Qwen3.8-27B-NVFP4</a>
Run the unsloth if you are using llama.cpp (GGUF)<p>Run the one you linked if you are running vllm (safetensors)
Considering the clusterfuck that is opus 5 or even fable, if Qwen 27B is trully better than Opus 4.7 Max, I will rejoice.
That's crazy, considering the massive size difference. But the small Qwen models are known for punching above their weight.
I wish each quant was benchmarked on the same tests as the original network so we could compare their performance
Unsloth publishes KL divergence numbers which measures how much the quantised probability distribution changes vs unquantised: <a href="https://unsloth.ai/docs/models/qwen3.8#quantization-analysis">https://unsloth.ai/docs/models/qwen3.8#quantization-analysis</a><p>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.
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.
> The fact that Unsloth only just started publishing KL divergences shows how unserious the quantization space is.<p>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.
That's not a replacement for benchmarks
Good morning Dario!
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.<p>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.<p>I've tried most models except Grok.<p>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.<p>Claude is slow and expensive (the cache hit prices are absurd).<p>OAI is pretty good, I might add it to my arsenal seeing how cheap it is.<p>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).<p>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.
You should check out Grok, it's quite a good deal from the Cursor subscription side but it's cheap even by API prices.
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.<p>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've settled on GLM-5.3 (formerly Deepseek v4 pro 0813) for architecting<p>Dude, GLM-5.3 released _today_.<p>The phrasing "I've settled on" is incorrect for this context.
i think you will like luna if you haven't tried it yet
As usual, the Jinja templates are messed up so use this [0] to reduce or turn off thinking, fix tool calling, keep a 100% KV cache hit rate, etc.<p>[0] <a href="https://huggingface.co/froggeric/Qwen-Fixed-Chat-Templates" rel="nofollow">https://huggingface.co/froggeric/Qwen-Fixed-Chat-Templates</a>
<a href="https://xcancel.com/Alibaba_Qwen/status/2088280182356611304" rel="nofollow">https://xcancel.com/Alibaba_Qwen/status/2088280182356611304</a>
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.<p>I'm sure there's some benchmaxxing going on, and some things you get only with a a larger model.<p>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.
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)<p>I saw unsloth has 1bit quants too so I might check that out, anybody have experience with those?
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.<p>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.<p>Also, I haven't used it enough yet to see if it's prone to infinite looping, but its predecessors sure were.
Reduce or turn off thinking:<p><a href="https://huggingface.co/froggeric/Qwen-Fixed-Chat-Templates" rel="nofollow">https://huggingface.co/froggeric/Qwen-Fixed-Chat-Templates</a>
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.
gemma4 can't hold a candle to 3.6
You can add a thinking budget thats not much effort in llamacpp. You can align the cut off message with your agent instructions.<p>What you describe is a engineering harness problem.<p>If you, and i mean the royal you, actually read tge thinking traces you can see and figure out where its stuck<p>This means an effective harness would observe when the model is overthinking and step in with reasonable redirection, like increasing logging.<p>Llamacpp can set reasoning budget and message per reauest, so it can be dynamic.<p>Your complaint is "skill issue" based and will be resolved by people who do something ither than vibe code react demos.
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?
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
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.
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.<p>Unsloth: <a href="https://huggingface.co/unsloth/Qwen3.8-27B-GGUF" rel="nofollow">https://huggingface.co/unsloth/Qwen3.8-27B-GGUF</a><p>This might work for you, but I didn't get on very well with MTPLX when I tried it a while back; YMMV: <a href="https://huggingface.co/Youssofal/Qwen3.8-27B-MTPLX-Optimized-Speed" rel="nofollow">https://huggingface.co/Youssofal/Qwen3.8-27B-MTPLX-Optimized...</a>
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).<p>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.<p>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.<p>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.
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.
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).
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:<p>2 bpw - ~85.7t/s<p>3 bpw - ~58.0t/s<p>4 bpw - ~43.9t/s<p>6 bpw - ~29.5t/s<p>8 bpw - ~22.2t/s<p>16 bpw - ~11.2t/s<p>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.<p>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.<p>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).<p>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.
Check out MTPLX and limit your context size.
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.<p>Insane if that is the case. Downloading now!
Unsloth Q4_K_M on a single 3090, llama.cpp
"Generate an SVG of a pelican riding a bicycle" first try
<a href="https://www.reddit.com/r/LocalLLaMA/comments/1voa3ch/comment/p3o0om9/" rel="nofollow">https://www.reddit.com/r/LocalLLaMA/comments/1voa3ch/comment...</a>
The file "Just loads" on llama.cpp, the Unsloth <a href="https://huggingface.co/unsloth/Qwen3.8-27B-GGUF" rel="nofollow">https://huggingface.co/unsloth/Qwen3.8-27B-GGUF</a> 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.<p>Most of the benchmark improvements afaict are in agentic and instruction following benchmarks.
This is one of the most important model releases since most use cases don't need SOTA/Frontier<p>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 <a href="https://x.com/ErdalToprak/status/2088299678085308761?s=20" rel="nofollow">https://x.com/ErdalToprak/status/2088299678085308761?s=20</a>
27b dense model at Opus 4.6 level<p>Opus at home<p>I hope there also will be a new ~10b variant
Qwen 3.6 is ~$2/m tok, 3.8 should be drop in replacement. Gemma 31B is $0.34/m tok. The price differential on these models is massive on openrouter.
[delayed]
Where do you see that? From what I can see on Open Router, Qwen 3.6 27B (the closest dense equivalent to Gemma 31) is $0.28/m. Am I missing something?<p><a href="https://openrouter.ai/qwen/qwen3.6-27b" rel="nofollow">https://openrouter.ai/qwen/qwen3.6-27b</a>
Man what a week,
We just had GLM 5.3 that came out and then we had smaller local model Qwen3.8-27B from Qwen<p>Just tried using Pi Agent and looks very promising
These are massive improvements - and something you can actually run on a laptop.
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.<p>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'm wondering whether any provider can offer this for cheaper $/token than the new DSv4 Flash, which is both cheaper and smarter :/<p>Completely local use is a different story, of course.
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) ?<p>Perf improvements seem to all come from training?
Wow. Speed improved as well. 200t/s on a RTX 5090!<p><a href="https://x.com/sgl_project/status/2088281320422322413" rel="nofollow">https://x.com/sgl_project/status/2088281320422322413</a>
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?
Was recently thinking about doing something similar, do you basically just have the qwens describe what they see for the flashes?<p>Was considering adding a LoRa/vision head to Flash, but seems like it could take a while to get it right.<p>If DSv4 Flash was multimodal, I’d probably be done model shopping for a while
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.<p>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.
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 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.
Others in this thread said it runs on RTX 4090.
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.
Really excited to see what people do with this. 3.7 27B was probably the best compromise between size and intelligence to run on consumer hardware
Why weren't the points merged again from the "dupe" thread that had 289 points?<p><a href="https://news.ycombinator.com/item?id=49299684">https://news.ycombinator.com/item?id=49299684</a><p>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.<p>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 am so happy right now, qwen3.6-27b was an absolute game changer. To see another one in the same league.. phew
Note: on the model card the comparison to Opus is Opus 4.6 Max, not 4.7
Looks like a pretty significant improvement on the DeepSWE benchmark compared to the previous 27B model.
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 ?<p>I only trust those users genuine personal tests
q4km is about 48 tps on a 4090. my llama.cpp params are --flash-attn on --parallel 1 --load-mode mmap
With spec decode should easily get to >100tps<p>on my dual 3090s qwen 3.5 27b was running at around 110tps using the config from <a href="https://github.com/noonghunna/club-3090" rel="nofollow">https://github.com/noonghunna/club-3090</a><p>make that 200tps on a single 5090, 4x faster than opus <a href="https://x.com/radixark/status/2088285681131110446" rel="nofollow">https://x.com/radixark/status/2088285681131110446</a><p>devs about to get handed a two 5090 box each and told to max that out
Can anyone confirm whether this new Qwen release is any more concise when thinking? Overthinking was the biggest (only?) downside of the Qwen models.
3.6-27B on little-coder was already mind blowing. looking forward to this guy!
Monstrous benchmarks! Hoping it is not benchmaxxed.
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.
also cool: Qwen 3.8 27b is multi modal!
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).
Seems to be SOTA for its size. Hopefully independent benchmarks will come soon.
Welcome deepseek flash flash!
Kinda was expecting to see Gemma 4 26B in benchmark comparisons :(
<a href="https://news.ycombinator.com/item?id=48403639">https://news.ycombinator.com/item?id=48403639</a><p>my prediction was way too far out. 4.6 at home! Woo.
remember to let llama.cpp catch up to anything new in this model. Save your judgment until about 2 weeks of use.
This with ddg mcp to fill in world knowledge. Are local models the future when computer architectures catch up?
NICE, i've been waiting for this drop, thanks for posting this
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...<p><i>" 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.<p>Let me start by reading the main project documentation and exploring the directory structure.<p>I'll take a look around this repo. Let me start by getting a lay of the land.<p>read resource CLAUDE.md (ctrl+o to expand)<p>ENOENT: no such file or directory, access 'XXXX/CLAUDE.md'"</i>
This is the homelab model hands down
My Strix Halo is about to go overdrive!
need another fable uncensored merge with 3.8, really curious what it can deliver
[flagged]
[flagged]
[flagged]