I don't think I can handle another small model release by qwen, I'm still trying to find the limits of 3.6 27B and they are already threatening us with a new one?
But jokes aside, I love the fast iteration, these are most probably again finetunes on the 3.5 architecture that appear better in internal testing, which is still very nice to see. Putting more and more pressure on the bigger labs to perform better is always a good thing.
Finetuning takes little resources, the base model training is the slow and expensive part. Architecturally 3.5 models are identical to their 3.6 counterparts, that is why there is a consensus that those are probably finetunes and not re-trained from scratch, like you will se many people publish their own on huggingface.
Understood, but look at their larger cadence over the years and the breadth of models. They are clearly not all finetunes. Meta for all its billions, doesn't have anything comparable.
In the china AI scene, there seem to be two separate types of companies.
Companies or labs like deepseek that produce less but larger and more innovative models, so seem to be more research oriented.
then there are companies like z.ai (GLM), Minimax, and Qwen which focus more on commercializing the AI and so produce far more versions, but with far less improvements between them (usually fine tunes)
Commercial providers like anthropic probably do the same thing, maybe even without labeling it like a different version if the model is similiar enough.
> Meta for all its billions, doesn't have anything comparable.
Maybe nothing released to the public. I don't know that all of their models are public. I think all they really care about is that they aren't relying on one or two cloud providers for a critical piece of their infrastructure.
Can someone explain what the current state of model benchmarking is? If you try to look up what the best locally runnable model is, you get a bunch of random blog posts using idiosyncratic criteria to rank things seemingly based on one dude's opinion.
Ideally I would love to see a leaderboard with relatively objective ranking criteria that 1. lets you filter by open weight / locally runnable, 2. filter by date of release (nothing older than x), and 3. is agnostic to hardware requirements. I just want to know what the best model is. Let me worry about how I will afford to run it.
I love the llmfit project for seeing what will run on your hardware, but it would be nice to know what I'm missing out on by not having better hardware, thus why objective hardware-agnostic ratings would be helpful.
That would be nice, but it's not going to be possible.
Any open benchmark has a very short life, since it will be pulled in and DPO / RL trained quickly for benchmaxxing purposes. So, you'll need a private test to have a hope of something fair. (These also get leaked over time, btw, so even then there's a window of usability).
These are expensive to run.
Now consider that there might be 15-20 viable quants for a given open model release; someone would have to want to pay for these private evals to be run on them. Even then, a good read through unsloth's commits and blog posts will remind you that there's quite a lot of engineering work to be done to get model inference working properly, even for models released by frontier or near-frontier labs. So, you'd want to make sure that you have a replicable 'best engineered' deployment to evaluate, or at least one that's closest to your hardware and fits the bill.
Upshot - it's much faster to download and try out a model, and possibly cheaper too. Well, cheaper since hugging face is paying the bandwidth bills.
there are benchmarks that have nothing to do with the training material, but with how the models are capable of things like reading code: https://needle-bench.cc/
Generally, you give them a document and you ask them to retrieve some subsection of the document then rate them on what they retrieved.
You can always find enough random documents, or create your own, to always run these and you can make it arbitrarily long. It's definitely a valid non-maxxable context test.
This seems like a viable eval strategy. Presumably finding a bug requires some degree of understanding of the code, beyond just information retrieval.
However it probably does not measure things like prompt adherence or ability to create code that implements a specification?
>I just want to know what the best model is. Let me worry about how I will afford to run it.
This is a very typical manager question that I suppose many people have who fail to see the simple truth: There is no "best" model. There are only best models for certain use-cases. Sometimes you'll find these in custom community leaderboards on platforms like huggingface, but for most business applications you'll probably have to come up with your own benchmark. Most common benchmarks are pretty worthless by now because all the usual ones are being gamed hard by model providers, to the point that there are now sometimes drastic differences between models that perform very similarly on common benchmarks.
The best thing I have come up with is just make a bunch of prompts / tasks that I personally care about and need a model to know how to do. As an example, when qwen3.6 27B dropped, I ran it, kimi, claude and glm 5/5.1 on a bunch of LLM-architecture specific tasks (stuff like 'implement an incremental KV-cache for autoregressive transformer inference' or 'implement flash Attention backward pass with D-optimization') and analyze the results, who made tests, are the tests valid, does their implementation actually work or are they only claiming it to, that sort of thing.
It is a day/weekend worth of work, but I think this is the best way to determine if the model fits your need specifically. This is what lead me to finding out that qwen 27b outperformed even kimi on those tasks, and that opus tries gaslighting me when I give it a spec of something that has been proven, but no published solution exists online. All other models gave their best shot at solving it, opus just said it's not possible (even when I gave it the finished working product that obviously works).
Especially for small models (but also big ones) I think the only way to know if a model will improve your workflow is this, personal benchmarks, expanded over time, ran in private.
> Ideally I would love to see a leaderboard with relatively objective ranking criteria that 1. lets you filter by open weight / locally runnable, 2. filter by date of release (nothing older than x), and 3. is agnostic to hardware requirements. I just want to know what the best model is. Let me worry about how I will afford to run it.
Stick to artificialanalysis.ai it has become the norm
I am very interested in seeing new qwen models. Qwen3.6 27b is the first one that can do things and doesnt constantly loose "it's mind" and that can be run on a 3090 with a good context size. But it's sometimes getting into a loop.
I sort of thought this about qwen3.5 35b, finally a local model that isn't a complete waste of electricity, but "upgrading" to 3.6 35b left me disappointed. It seemed more like a downgrade. But honestly I've barely used either. Subjectively they still seem far from the frontier models, but for what they can do, it's great to be able to do locally.
Pretty decent, it's given similar book recommendations as Claude when I feed it my list of read books and thoughts on them. You'll have to tell them to never use emojis. I was using 3.5 a while ago to generate some flavor text while I was playing a bit of an old-school dungeon-crawler game (it's like Wizardry), a genre I don't particularly enjoy much, but it's funner with the flavor text. Worth setting up something like open webui or other front-ends since a pure CLI experience via ollama is pretty bad.
I had a flavor of an older version of Qwen (I forget which one to be fair) that was coding along, then lost itself in a loop, I was so confused, it was just a random greenfield "lets see how it does" type of project anyway.
Agreed. Incidentally, in my testing, qwen models (qwen3.6-35b-a3b and earlier 3.5) are WAY better with vision than gemma4-26b-a4b. I would normally want to stick with gemma4 only (I use it for spam filtering), but it just doesn't cut it for vision work, and qwen models do.
Qwen 3.5/3.6 are far better at vision. Even the 9B model beats Gemma 4 31B in my use case. They describe the scene more accurately and they focus on the important elements like a human would.
Gemma 4 frequently misses important element, doesn't understand what things are, and is very coy even if you ask for lots of detail. You have to give it hints "hey what's that round thing on the left" to get half decent answers.
(Yes I did set the min-tokens correctly. I also tested bf16 and Q8 to make sure it wasn't a quant issue.)
It's unfortunate because Gemma 4 is so so so much better at natural language interactions.
I have my own benchmark: a photo of filament boxes in my drawer. The task is to list (inventory) the filaments. It's not an easy task: the letters are small, the boxes are in various orientations, the labels have lots of writing on them, etc.
This is what qwen produces (basically perfect):
Fiberlogy (Blue Boxes - Left Side)
ASA Black: 750 g (0.75 kg)
Easy PLA Graphite: 850 g (0.85 kg)
Ref Easy PLA Gray: 850 g (0.85 kg)
Ref Easy PLA White: 850 g (0.85 kg)
Ref Easy PLA Inox: 850 g (0.85 kg)
Bambu Lab (Brown Boxes - Right Side)
ASA White: 750 g (0.75 kg)
PLA-CF (Burgundy Red): 1,000 g (1 kg)
PAHT-CF (Black): 500 g (0.5 kg)
PET-CF (Black): 500 g (0.5 kg)
PETG-CF (Black): 1,000 g (1 kg)
----------
And this is gemma-4-26b-a4b (basically useless):
### *Fiberlogy Filaments*
All Fiberlogy items listed below share these common specifications: Made in Poland, $1.75\text{mm}$ Diameter, $0.75\text{kg}$ Weight, and Print Temperature of $200-230^{\circ}\text{C}$.
* *PLA Black* (Barcode: 5903281997986)
* *Easy PLA Graphite* (Barcode: 5903281994057)
* *Ref Easy PLA Grey*
* *Ref Easy PLA White* (Barcode: 5903281994386)
* *Ref Easy PLA True...* (Text is cut off)
**
### *Bambu Lab Filaments*
All Bambu Lab items listed below share these common specifications: Made in China, Removable Spool (Do Not Remove).
* *PLA-CF (Carbon Fiber Reinforced)*
\* Color: Burgundy Red
\* Diameter: $1.75 \pm 0.02\text{mm}$
\* Weight: $1.0\text{kg}$
\* Suggested Drying Conditions: $45^{\circ}\text{C}$ for $6-12$ hours
* *PAHT-CF (High Temperature Polyamide with Carbon Fiber)*
\* Color: Black
\* Diameter: $1.75 \pm 0.02\text{mm}$
\* Weight: $0.5\text{kg}$
\* Suggested Drying Conditions: $80^{\circ}\text{C}$ for $6-12$ hours
Thanks. Did you set the image min/max tokens for Gemma4 to 1120 for this? This might not be a fair comparison without that, to the differences in architecture.
I did not, and I had no idea such a setting even existed. This could definitely change things. However, I don't see a way to set this in LM Studio, which is what I currently use to run models.
I'm not much interested in vibe coding (for those who aren't aware that LLMs have other uses). The specific model I've been using with Ollama is hf.co/unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF:UD-Q4_K_XL and it's amazing how fast it is on 64 GB of RAM and i5-13400 CPU. No GPU on this computer. Gemma 4 E4B will think for a couple of minutes vs 3-5 seconds for Qwen. It's hard to believe how much you can do with such limited hardware using their models.
Qwen 3.6 35B (finetuned) is so good that it became standard open weights for everyday use. Is not far at all from proprietary models if you give it tools, skills and agents etc, it can actually finish the job. (Thank you Qwen team, appreciated). Using opensource now we can definitely rely to design from scratch very complicated architecture and build pretty fast the full pack.
Wish to see Europe AI unleashed, wake up.
> Is not far at all from proprietary models if you give it tools, skills and agents etc,
I use Qwen 3.6 27B, the dense version of this model which is slightly better.
I don't agree that it's close at all. Maybe for some small, easy tasks, but not for working on real codebases. It's amazing for something I can run at home, but the difference between it and Opus or GPT-5.5 is huge.
Really, how so? Because we work with codebases daily, can you tell us a concrete example!
In our case we work in consumer hardware (ish), 10 million ctx (1 million output, 1 million input proven, sometimes it loops or breaks at over 500k ctx byt at ~17tps linear). IT can read the full codebase, unleash agents, and write in disk editing and patching files creating a full app in 3-4 minutes. IT can do Web search and Rag pretty fast, it understands and fix the user query, sys prompts and adapt/fix them if needed on the fly. I am wondering what more do you do?
Edit: Forgot to mention that it can process images and pdf, and 100s of other files, it can even create presentations in code or mermaid, svg, charts js etc.
Here a basic version of it: https://hugston.com/chat
how do you do 1mio context with qwen3.6 27b, that only supports 256k? and what hardware would you run that on? 2 * 3090 is afaik currently at max 256k context.
You can get all the Qwen 3.x models up to ~1 million tokens using YaRN with llama.cpp.[0]
Personally I am using `--no-context-shift` and feeding in context back in on failure at the harness level.
I have 2x1080ti + 1xTitanV that have a full 262,144 tokens context on 262,144 tokens with `-sm tensor` at 62.04 t/s which isn't so bad.
But I also have a 1x3090 running unsloth/Qwen3.6-27B-MTP-GGUF:UD-Q4_K_XL at 41.89 t/s but with only 130k context, but if you have a modular programming style both work pretty well.
podman build -t local/llama.cpp:full-cuda --target full -f .devops/cuda.Dockerfile .
And here is the logs from a 'make me a flappy bird program in python' webui prompt.
prompt eval time = 105.86 ms / 19 tokens ( 5.57 ms per token, 179.47 tokens per second)
eval time = 100549.41 ms / 4608 tokens ( 21.82 ms per token, 45.83 tokens per second)
total time = 100655.28 ms / 4627 tokens
draft acceptance rate = 0.47215 ( 3408 accepted / 7218 generated)
That config looked too complicated, getting rid of the --prio 3 and --poll 100, setting the draft-n-max to now recommended values, etc... kicked it up to 61 t/s
You can increase the context window beyond its max trained context using RoPE scaling[0] which will require more VRAM.
But you can increase your context window for the same VRAM by quantizing the KV cache with FP8 (double the context) or TurboQuant (more than double)[1].
As funny as it may sound a q4_k_m well converted and quantized properly (and finetuned, impereative) would do the job. The 27b it may be good but is heavy, it burns the hardware. I personally prefer the 397B if I am stucked and can´t progress, it can still run with 7 tps. Now with the Mtp (multitoken prediction) it nearly double the speed ( reached 82tps today with the 35b 100000ctx). I recommend it you give it a try.
You don't pick just one model to "work on real codebases". You use a very advanced model to plan, and a not-very-advanced, cheaper, faster model to execute planned tasks. This saves money and speeds up work. This is the guidance from Anthropic & OpenAI.
Bad is mystifying. Unassisted but for handing it a pile of PDFs of relevant academic papers and my initial codebase, I had hermes agent based on qwen-3.6 27B implement karatsuba multiplication of characteristic-2 polynomials in C++ in an existing codebase with an internal field arithmetic library. It correctly found the 'obvious' optimizations using the field properties. Then I had it implement the recursive halfgcd algorithm for these polynomials using it.
It wrote extensive test cases and validated them with mutation testing (per my standard instructions)-- took many tries getting the algorithms right but with the tests handy it found and fixed the errors.
Here is a dataset you can choose from: https://huggingface.co/datasets/Avtrkrb/combined-reasoning-o...
Get a 10000 samples from it according to your needs and go for it. The key (in my opinion) is not cutting the Sequence Length among other things. Whatever traditional finetuning repo will do, if your hardware supports it Unsloth is faster.
I love that open weight models are catching up so quickly. Also hilarious how far behind Grok is. I guess demand for Grok must be poor if Anthropic is able to rent resources from xAI.
Just to be clear, "Plus" and "Max" Qwen models are closed. Seems likely smaller open versions will be released, but that's not what was announced today
To play devil's advocate I do feel like Grok has a unique "feel" to it. All the Chinese models feel like GPT or Claude distillations, but Grok has a certain unique way of saying and doing things. But that said, it also feels a year behind the state of the art.
Right now they want to prevent the US labs from gaining any sort of self-reinforcing oligopoly on the space, and to let the ecosystem in China flourish.
Will they release the large models as open weight too? So far it seems only 35 or 27 B etc models are being released with nothing larger unlike before.
Gemma 4 and Qwen 3.6 were when my local inference experiments graduated from toy challenges with much hand holding to actually full day back and forth with good ability to utilise tool calls to discover how things are glued together.
I'm not talking about greenfield dev, I'm talking about interfacing with an existing decade old codebase.
I stopped caring about benchmarks at MiniMax M2.5. I no longer want more advanced models. I want cheaper models that don't slow down when everyone else is online.
I have a tangential question. Provided that it is correct that current proprietary models are offered at below cost-covering rates (I believe this is a consensus if I'm not mistaken¹); what factor (multiplication) would have to be applied approximately to current rates to reach break even?
¹: I think I read this a couple of times but I'm not sure if correct to begin with. Can this be substantiated based on annual financial reporting or other published business metrics by OpenAI, Anthropic et al.?
I don’t mind a principled downvote, but can a downvoter explain his or her reasoning? Genuinely curious. I found the linked rankings surprising, and think of myself as relatively well informed. Please, enlighten me..
But jokes aside, I love the fast iteration, these are most probably again finetunes on the 3.5 architecture that appear better in internal testing, which is still very nice to see. Putting more and more pressure on the bigger labs to perform better is always a good thing.
reply