Rendered at 17:20:53 GMT+0000 (Coordinated Universal Time) with Cloudflare Workers.
hypfer 48 minutes ago [-]
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
Aurornis 17 minutes ago [-]
> --cache-type-k q8_0 --cache-type-v q8_0
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.
hypfer 14 minutes ago [-]
Yes to both.
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 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.
reilly3000 37 minutes ago [-]
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!
hypfer 35 minutes ago [-]
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
nubg 14 minutes ago [-]
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.
williamcotton 4 minutes ago [-]
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.
You sound like your trying to reassure yourself of something.
I sure hope my boss doesn't think he built my work! He'd probably get fired pretty quickly during on call!
3 minutes ago [-]
D4Ha 28 minutes ago [-]
Do you find it useful or worthwhile to split a large LLM across two GPUs on a desktop?
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?
bilekas 1 minutes ago [-]
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?
mips_avatar 2 minutes ago [-]
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.
hypfer 27 minutes ago [-]
I just have one 4090 and nothing else, soooo I can't help there, but maybe someone else can.
CamperBob2 20 minutes ago [-]
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.
bmitc 27 minutes ago [-]
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.
Aurornis 11 minutes ago [-]
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.
hypfer 25 minutes ago [-]
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.
porphyra 18 minutes ago [-]
> 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.
manquer 13 minutes ago [-]
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.
solenoid0937 14 minutes ago [-]
Sure, but front ends like LMStudio exist for that crowd
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
bmitc 15 minutes ago [-]
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.
skrebbel 16 minutes ago [-]
> LM Studio doesn't work behind proxies.
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.
bmitc 14 minutes ago [-]
> 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.
skrebbel 10 minutes ago [-]
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.
xienze 26 minutes ago [-]
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.
Auracle 22 minutes ago [-]
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”?
You could override, obviously.
zargon 16 minutes ago [-]
Yes, llama.cpp does that.
naasking 18 minutes ago [-]
You know free LLMs can help you understand that command line or design your own...
CamperBob2 22 minutes ago [-]
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.
KronisLV 2 hours ago [-]
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.
Phemist 30 minutes ago [-]
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.
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.
peri-cl 2 hours ago [-]
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.
expedited123 37 minutes ago [-]
Mind sharing your laptops specs? Just interested to see what is needed to locally run Qwen3.6-35B-A3B
peri-cl 9 minutes ago [-]
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 that's >140k tokens.
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).
Alifatisk 2 hours ago [-]
> I'd gladly take A5B or A8B or even A10B as a sort of middle ground.
Whats up with focusing on the active param count? Do yall fiddle with the weights or something?
kennywinker 52 minutes ago [-]
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
Alifatisk 28 minutes ago [-]
Thank you Kenny
martinald 2 hours ago [-]
You can run these on CPUs at a somewhat reasonable speed.
KronisLV 1 hours ago [-]
Or (somewhat) low TDP GPUs for that matter, like workstation ones, that might have enough total VRAM but not the best bandwidth/compute.
scrlk 2 hours ago [-]
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.
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.
spmurrayzzz 2 hours ago [-]
> They do not beat opus on real-world usage
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.
croemer 22 minutes ago [-]
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.
cyanydeez 44 minutes ago [-]
Let us know when you have Qwen vs Qwen comparison stats. As long as there's not a regression, that'd be awesome.
spmurrayzzz 41 minutes ago [-]
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.
croemer 21 minutes ago [-]
Depends on where you sit on the binomial curve. At p=0.04 for n=250 4% points would not be within margin of error.
KronisLV 2 hours ago [-]
> ...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.
pimeys 1 hours ago [-]
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.
niek_pas 2 hours ago [-]
A wise man once said, "not everything that counts can be counted, and not everything that can be counted counts".
mlmonkey 1 hours ago [-]
In the end, the only benchmark that matters is your own.
bewareofscams 2 hours ago [-]
Only useful benchmarks are those you (in particular) don't have access to.
rhdunn 28 minutes ago [-]
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.
There are tools like promptfoo designed for this.
xienze 2 hours ago [-]
> 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."
redox99 23 minutes ago [-]
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.
metadat 47 minutes ago [-]
How can you say this when you haven't even tried it yet? Is it just hypothetical vibes?
willcmcc 53 minutes ago [-]
There is 0 shot you can make that claim about this model you have not used or downloaded yet
altmanaltman 42 minutes ago [-]
"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.
jrflo 22 minutes ago [-]
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.
CuriouslyC 3 minutes ago [-]
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.
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.
benxh 2 hours ago [-]
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
petu 2 hours ago [-]
Unsloth one is gguf for llama.cpp (and some other on-device engines).
So advantage is not having to produce your own quantisation / gguf from .safetensors you've linked.
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.
zargon 44 minutes ago [-]
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.
lostmsu 19 minutes ago [-]
> 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.
WithinReason 19 minutes ago [-]
That's not a replacement for benchmarks
UncleOxidant 1 hours ago [-]
Good morning Dario!
ramon156 2 hours ago [-]
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.
satvikpendem 8 minutes ago [-]
You should check out Grok, it's quite a good deal from the Cursor subscription side but it's cheap even by API prices.
ai_fry_ur_brain 2 minutes ago [-]
[dead]
simplyluke 1 hours ago [-]
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.
hypfer 2 hours ago [-]
> I've settled on GLM-5.3 (formerly Deepseek v4 pro 0813) for architecting
Dude, GLM-5.3 released _today_.
The phrasing "I've settled on" is incorrect for this context.
ramon156 2 hours ago [-]
hence the "former deepseek v4 pro". I tried it out this morning and have had no complaints. I already liked glm 5.2
hypfer 2 hours ago [-]
The sentence still doesn't make sense, because "settled on" implies a long testing phase with a verdict eventually emerging out of that.
What you're currently doing is "testing out"
Topfi 39 minutes ago [-]
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…
satvikpendem 9 minutes ago [-]
What are you working on? That can dictate which models are best.
tosh 1 hours ago [-]
i think you will like luna if you haven't tried it yet
satvikpendem 15 minutes ago [-]
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.
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?
Der_Einzige 5 minutes ago [-]
Yes yes, oh god yes. They also spread FUD in the form of terrible recommended sampler settings.
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.
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.
DanielHB 2 hours ago [-]
What kind of things you only get with a larger model?
redox99 14 minutes ago [-]
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.
onlyrealcuzzo 33 minutes ago [-]
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.
alex7o 31 minutes ago [-]
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.
versteegen 2 hours ago [-]
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.
jedbrooke 2 hours ago [-]
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?
spwa4 56 minutes ago [-]
Sounds like you need to check what the max context is set to ...
jedbrooke 41 minutes ago [-]
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
Casteil 1 hours ago [-]
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.
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.
ThouYS 1 hours ago [-]
gemma4 can't hold a candle to 3.6
cyanydeez 1 hours ago [-]
You can add a thinking budget thats not much effort in llamacpp. You can align the cut off message with your agent instructions.
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.
davidw 1 minutes ago [-]
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?
Almondsetat 29 minutes ago [-]
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 $5-10k for running DeepSeek this is your best budget option for local refactor/small scale dev help
LeBit 1 minutes ago [-]
I understand the B70 is a bargain vs AMD and especially nVidia offerings, but to me it feels like I would be buying something that would feel too limited in less than a year. 48G would be much more confortable.
And I know the 96G nVidia cards are selling for over 10k$.
The future can’t arrive fast enough!
bogzz 27 minutes ago [-]
Oh, can it work with the /v1/completions/ auto-complete endpoint?
Almondsetat 21 minutes ago [-]
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
bogzz 19 minutes ago [-]
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.
TomGarden 2 hours ago [-]
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.
mft_ 2 hours ago [-]
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.
jwr 33 minutes ago [-]
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.
seanmcdirmid 39 minutes ago [-]
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.
UncleOxidant 1 hours ago [-]
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).
anana_ 1 hours ago [-]
Seems like MTP is available immediately!
LoganDark 2 hours ago [-]
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.
brcmthrowaway 2 hours ago [-]
Check out MTPLX and limit your context size.
NorwegianDude 2 hours ago [-]
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.
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.
minimaltom 28 minutes ago [-]
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.
erdaltoprak 2 hours ago [-]
This is one of the most important model releases since most use cases don't need SOTA/Frontier
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
mring33621 1 hours ago [-]
9B Qwen models are good and fast for local python coding tasks.
tosh 2 hours ago [-]
they are all overlapping but:
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)
32 minutes ago [-]
piyh 2 hours ago [-]
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.
SparkyMcUnicorn 1 minutes ago [-]
[delayed]
jjice 49 minutes ago [-]
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?
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
Just tried using Pi Agent and looks very promising
chvid 2 hours ago [-]
These are massive improvements - and something you can actually run on a laptop.
mickeyp 2 hours ago [-]
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.
theanonymousone 2 hours ago [-]
I'm wondering whether any provider can offer this for cheaper $/token than the new DSv4 Flash, which is both cheaper and smarter :/
Completely local use is a different story, of course.
minimaltom 1 hours ago [-]
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) ?
Perf improvements seem to all come from training?
anana_ 1 hours ago [-]
As was the case with GLM 5.3, it seems that there is still much juice to be squeezed from post-training
bertili 1 hours ago [-]
Wow. Speed improved as well. 200t/s on a RTX 5090!
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?
ericd 55 minutes ago [-]
Was recently thinking about doing something similar, do you basically just have the qwens describe what they see for the flashes?
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
arjie 28 minutes ago [-]
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.
irthomasthomas 1 hours ago [-]
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.
spwa4 55 minutes ago [-]
Pretty sure you can use Gemma models on Google's "Vertex AI".
synergy20 2 hours ago [-]
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.
KyleJune 11 minutes ago [-]
Others in this thread said it runs on RTX 4090.
baron3dl 2 hours ago [-]
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.
TomGarden 2 hours ago [-]
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
ThouYS 2 hours ago [-]
I am so happy right now, qwen3.6-27b was an absolute game changer. To see another one in the same league.. phew
lossolo 11 minutes ago [-]
Why weren't the points merged again from the "dupe" thread that had 289 points?
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.
jlkivey 2 hours ago [-]
Note: on the model card the comparison to Opus is Opus 4.6 Max, not 4.7
2 hours ago [-]
kunver 2 hours ago [-]
Looks like a pretty significant improvement on the DeepSWE benchmark compared to the previous 27B model.
yassa9 2 hours ago [-]
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 ?
I only trust those users genuine personal tests
alyandon 2 hours ago [-]
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.
devs about to get handed a two 5090 box each and told to max that out
naasking 13 minutes ago [-]
Can anyone confirm whether this new Qwen release is any more concise when thinking? Overthinking was the biggest (only?) downside of the Qwen models.
ThouYS 2 hours ago [-]
3.6-27B on little-coder was already mind blowing. looking forward to this guy!
anana_ 2 hours ago [-]
Monstrous benchmarks! Hoping it is not benchmaxxed.
tristor 22 minutes ago [-]
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.
tosh 2 hours ago [-]
also cool: Qwen 3.8 27b is multi modal!
gurkwart 2 hours ago [-]
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).
pu_pe 2 hours ago [-]
Seems to be SOTA for its size. Hopefully independent benchmarks will come soon.
kunver 2 hours ago [-]
Welcome deepseek flash flash!
2 hours ago [-]
expedited123 2 hours ago [-]
Kinda was expecting to see Gemma 4 26B in benchmark comparisons :(
kamranjon 2 hours ago [-]
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.
my prediction was way too far out. 4.6 at home! Woo.
altruios 2 hours ago [-]
remember to let llama.cpp catch up to anything new in this model. Save your judgment until about 2 weeks of use.
chrismartin 2 hours ago [-]
'Good' news, there seems to be nothing new architecture-wise. Same as Qwen 3.5 and 3.6, so llama.cpp doesn't know the difference.
brcmthrowaway 2 hours ago [-]
This with ddg mcp to fill in world knowledge. Are local models the future when computer architectures catch up?
alpha_trion 2 hours ago [-]
NICE, i've been waiting for this drop, thanks for posting this
cmrdporcupine 34 minutes ago [-]
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'"
Mr_Eri_Atlov 2 hours ago [-]
This is the homelab model hands down
brcmthrowaway 2 hours ago [-]
My Strix Halo is about to go overdrive!
ramon156 2 hours ago [-]
need another fable uncensored merge with 3.8, really curious what it can deliver
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 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.
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.
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...
I sure hope my boss doesn't think he built my work! He'd probably get fired pretty quickly during on call!
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?
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.
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.
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.
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.
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
For Claude, I setting a single config file and then download and run Claude Code CLI. Even easier for the GitHub Copilot CLI.
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.
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.
You could override, obviously.
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.
Edit: link - https://huggingface.co/rico03/Qwen3.6-27B-Claude-Opus-Reason...
https://huggingface.co/froggeric/Qwen-Fixed-Chat-Templates
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 that's >140k tokens.
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).
Whats up with focusing on the active param count? Do yall fiddle with the weights or something?
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.
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.
There are tools like promptfoo designed for this.
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."
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.
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.
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.
Dude, GLM-5.3 released _today_.
The phrasing "I've settled on" is incorrect for this context.
What you're currently doing is "testing out"
[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?
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.
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.
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.
I saw unsloth has 1bit quants too so I might check that out, anybody have experience with those?
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
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.
https://huggingface.co/froggeric/Qwen-Fixed-Chat-Templates
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.
And I know the 96G nVidia cards are selling for over 10k$.
The future can’t arrive fast enough!
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...
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.
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.
Insane if that is the case. Downloading now!
Most of the benchmark improvements afaict are in agentic and instruction following benchmarks.
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
Opus at home
I hope there also will be a new ~10b variant
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)
https://openrouter.ai/qwen/qwen3.6-27b
Just tried using Pi Agent and looks very promising
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.
Completely local use is a different story, of course.
Perf improvements seem to all come from training?
https://x.com/sgl_project/status/2088281320422322413
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.
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.
I only trust those users genuine personal tests
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.
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
my prediction was way too far out. 4.6 at home! Woo.
" 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'"