56 comments

[ 3.0 ms ] story [ 72.7 ms ] thread
Are there any non-Chinese open models that offer comparable performance?
I think you could look into Minstral. There's also GPT-OSS but I'm not sure how well it stacks up.

What's your problem with Chinese LLMs?

All the western ones are closed while all the Chinese ones are open. The only exception is the European Mistral but performance of that model is not very satisfactory. Hopefully they make some improvements soon
What's the problem with Chinese models? The models are already open which makes them more trustworthy than the American closed models.
It's not only "non-Chinese" to think about here. There's nobody really touching Qwen in the single-GPU size class and there hasn't been for a couple of generations.
Why would anyone care if its Chinese? No one uses ChatGPT because its from the US.
The new 35b model is great. That said, it has slight incompatibility's with Claude Code. It is very good for tool use.
Claude code is designed for anthropic models. Try it with opencode!
Have you tried the 122B one?
What kind of hardware does HN recommend or like to run these models?
The cheapest option is two 3060 12G cards. You'll be able to fit the Q4 of the 27B or 35B with an okay context window.

If you want to spend twice as much for more speed, get a 3090/4090/5090.

If you want long context, get two of them.

If you have enough spare cash to buy a car, get an RTX Ada with 96G VRAM.

Radeon R9700 with 32 GB VRAM is relatively affordable for the amount of RAM and with llama.cpp it runs fast enough for most things. These are workstation cards with blower fans and they are LOUD. Otherwise if you have the money to burn get a 5090 for speeeed and relatively low noise, especially if you limit power usage.
I think the 27B dense model at full precision and 122B MoE at 4- or 6-bit quantization are legitimate killer apps for the 96 GB RTX 6000 Pro Blackwell, if the budget supports it.

I imagine any 24 GB card can run the lower quants at a reasonable rate, though, and those are still very good models.

Big fan of Qwen 3.5. It actually delivers on some of the hype that the previous wave of open models never lived up to.

It depends. How much are you willing to wait for an answer? Also, how far are you willing to push quantization, given the risk of degraded answers at more extreme quantization levels?
Macs or a strix halo. Unless you want to go lower than 8-bit quantization where any GPU with 24GBs of VRAM would probably run it.
For 27B, just get a used 3090 and hop on to r/LocalLLaMA. You can run a 4bpw quant at full context with Q8 KV cache.
Smells like hyperbole. A lot of people making such claims don’t seem to have continued real world experience with these models or seem to have very weird standards for what they consider usable.

Up until relatively recently, while people had already long been making these claims, it came with the asterisks of „oh, but you can’t practically use more than a few K tokens of context“.

Qwen3.5-122B-A10B BF16 GGUF = 224GB. The "80Gb VRAM" mentioned here will barely fit Q4_K_S (70GB), which will NOT perform as shown on benchmarks.

Quite misleading, really.

qwen 3.5 is really decent. oOtside for some weird failures on some scaffolding with seemingly different trained tools.

Strong vision and reasoning performance, and the 35-a3b model run s pretty ok on a 16gb GPU with some CPU layers.

I asked it to recite potato 100 times coz I wanted to benchmark speed of CPU vs GPU. It's on 150 line of planning. It recited the requested thing 4 times already and started drafting the 5th response.

...yeah I doubt it

I periodically try to run these models on my MBP M3 Max 128G (which I bought with a mind to run local AI). I have a certain deep research question (in a field that is deeply familiar to me) that I ask when I want to gauge model's knowledge.

So far Opus 4.6 and Gemini Pro are very satisfactory, producing great answers fairly fast. Gemini is very fast at 30-50 sec, Opus is very detailed and comes at about 2-3 minutes.

Today I ran the question against local qwen3.5:35b-a3b - it puffed for 45 (!) minutes, produced a very generic answer with errors, and made my laptop sound like it's going to take off any moment.

Wonder what am I doing wrong?.. How am I supposed to use this for any agentic coding on a large enough codebase? It will take days (and a 3M Peltor X5A) to produce anything useful.

Try with qwen 3.5 122b; it has more parameters so a larger corpus of knowledge to draw from than 35b.
Thinking about getting a new MBP M5 Max 128GB (assuming they are released next week). I know "future proofing" at this stage is near impossible, but for writing Rust code locally (likely using Qwen 3.5 for now on MLX), the AIs have convinced me this is probably my best choice for immediate with some level of longevity, while retaining portability (not strictly needed, but nice to have). Alternatively was considering RTX options or a mac studio, but was leaning towards apple for the unified memory. What does HN think?
this is bullshit with a kernel of truth.

none of the qwen 3.5 models are anywhere near sonnet 4.5 class, not even the largest 397b.

BUT 27b is the smartest local-sized model in the world by a wide wide margin. (35b is shit. fast shit, but shit.)

benchmarks are complete, publishing on Monday.

SWE chart is missing Claude on front page, interesting way to present your data. Mix and match at will. Grown up people showing public school level sneakiness. That fact alone disqualifies your LL. Business/marketing leaders usually are brighter than average developers... so there.
Can it do FizzBuzz in Brainfuck? Thus far all local models have tripped over their feet or looped out.
If you're new to this: All of the open source models are playing benchmark optimization games. Every new open weight model comes with promises of being as good as something SOTA from a few months ago then they always disappoint in actual use.

I've been playing with Qwen3-Coder-Next and the Qwen3.5 models since they were each released.

They are impressive, but they are not performing at Sonnet 4.5 level in my experience.

I have observed that they're configured to be very tenacious. If you can carefully constrain the goal with some tests they need to pass and frame it in a way to keep them on track, they will just keep trying things over and over. They'll "solve" a lot of these problems in the way that a broken clock is right twice a day, but there's a lot of fumbling to get there.

That said, they are impressive for open source models. It's amazing what you can do with self-hosted now. Just don't believe the hype that these are Sonnet 4.5 level models because you're going to be very disappointed once you get into anything complex.

Depends on what you expect from the model. For coding/agentic tasks there is SWE Bench https://www.swebench.com/ which gives a better picture. MiniMax, GLM and Kimi K2 seem to be better models for this purpose than Qwen. And it matches my (limited) actual experience.
How much computer do you need to make them work like Sonnet 4.5 from claude but locally?
Getting better, but definitely not there yet, nor near Sonnet 4.5 performance.

What these open models are great for are for narrow, constrained domains, with good input/output examples. I typically use them for things like prompt expansion, sentiment analysis, reformatting or re-arranging flow of code.

What I found they have trouble with is going from ambiguous description -> solved problem. Qwen 3.5 is certainly the best of the OSS models I've found (beating out GPT 120b OSS which was the previous king), and it's just starting to demonstrate true intelligence in unbound situations, but it isn't quite there yet. I have a RTX 6000 pro, so Qwen 3.5 is free for me to run, but I tend to default to Composer 1.5 if I want to be cheap.

The trend however is super encouraging. I bought my vid card with the full expectation that we'll have a locally running GPT 5.2 equiv by EoY, and I think we're on track.

A big part that a lot of local users forget is inference is hard. Maybe you have the wrong temperature. Maybe you have the wrong min P. Maybe you have the wrong template. Maybe the implementation in llama cpp has a bug. Maybe Q4 or even Q8 just won’t compare to BF16. Reality is, there’s so many knobs to LLM inferencing and any can make the experience worse. It’s not always the model’s fault.
No it does not. None of these models have the “depth” that the frontier models have across a variety of conversations, tasks and situations. Working with them is like playing snakes and ladders, you never know when it’s going to do something crazy and set you back.
Raw scale of parameters is POWER, you can't get performance out of a small model from a much larger one.