Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Qwen3.6-35B-A3B: Agentic coding power, now open to all

qwen.ai

Post reply on HN
Print view View original

21–30 of 563 posts

Re: Qwen3.6-35B-A3B: Agentic coding power, now open to all

#21

Post by julianlam »

NitpickLawyer wrote: post #13
fred_is_fred wrote: post #4

How does this compare to the commercial models like Sonnet 4.5 or GPT? Close enough that the price is right (free)?

> Close enough No. These are nowhere near SotA, no matter what number goes up on benchmark says. They are amazing for what they are (runnable on regular PCs), and you can find usecases for them (where privacy >> speed / accuracy) where they perform "good enough", but they are not magic. They have limitations, and you need to adapt your workflows to handle them.

Can you share more about what adaptations you made when using smaller models?

I'm just starting my exploration of these small models for coding on my 16GB machine (yeah, puny...) and am running into issues where the solution may very well be to reduce the scope of the problem set so the smaller model can handle it.

Re: Qwen3.6-35B-A3B: Agentic coding power, now open to all

#23

Post by Havoc »

bossyTeacher wrote: post #16

Does anyone have any experience with Qwen or any non-Western LLMs? It's hard to get a feel out there with all the doomerists and grifters shouting. Only thing I need is reasonable promise that my data won't be used for training or at least some of it won't. Being able to export conversations in bulk would be helpful.

The Chinese models are generally pretty good.

> Only thing I need is reasonable promise that my data won't be used

Only way is to run it local.

I personally don’t worry about this too much. Things like medical questions I tend to do against local models though

Re: Qwen3.6-35B-A3B: Agentic coding power, now open to all

#24

Post by mhitza »

abhikul0 wrote: post #7

I hope the other sizes are coming too(9B for me). Can't fit much context with this on a 36GB mac.

It's a MoE model and the A3B stands for 3 Billion active parameters, like the recent Gemma 4.

You can try to offload the experts on CPU with llama.cpp (--cpu-moe) and that should give you quite the extra context space, at a lower token generation speed.

Re: Qwen3.6-35B-A3B: Agentic coding power, now open to all

#25

Post by NitpickLawyer »

mtct88 wrote: post #9

Nice release from the Qwen team. Small openweight coding models are, imho, the way to go for custom agents tailored to the specific needs of dev shops that are restricted from accessing public models. I'm thinking about banking and healthcare sector development agencies, for example. It's a shame this remains a market largely overlooked by Western players, Mistral being the only one moving in that direction.

I agree with the sentiment, but these models aren't suited for that. You can run much bigger models on prem with ~100k of hardware, and those can actually be useful in real-world tasks. These small models are fun to play with, but are nowhere close to solving the needs of a dev shop working in healthcare or banking, sadly.

Re: Qwen3.6-35B-A3B: Agentic coding power, now open to all

#26

Post by nickthegreek »

pdyc wrote: post #22
abhikul0 wrote: post #7

I hope the other sizes are coming too(9B for me). Can't fit much context with this on a 36GB mac.

can you elaborate? you can use quantized version, would context still be an issue with it?

context is always an issue with local models and consumer hardware.

Re: Qwen3.6-35B-A3B: Agentic coding power, now open to all

#27

Post by adrian_b »

incomingpain wrote: post #2

Wowzers, we were worried Qwen was going to suffer having lost several high profile people on the team but that's a huge drop. It's better than 27b?

Their previous model Qwen3.5 was available in many sizes, from very small sizes intended for smartphones, to medium sizes like 27B and big sizes like 122B and 397B.

This model is the first that is provided with open weights from their newer family of models Qwen3.6.

Judging from its medium size, Qwen/Qwen3.6-35B-A3B is intended as a superior replacement of Qwen/Qwen3.5-27B.

It remains to be seen whether they will also publish in the future replacements for the bigger 122B and 397B models.

The older Qwen3.5 models can be also found in uncensored modifications. It also remains to be seen whether it will be easy to uncensor Qwen3.6, because for some recent models, like Kimi-K2.5, the methods used to remove censoring from older LLMs no longer worked.

Re: Qwen3.6-35B-A3B: Agentic coding power, now open to all

#28

Post by Mashimo »

bossyTeacher wrote: post #16

Does anyone have any experience with Qwen or any non-Western LLMs? It's hard to get a feel out there with all the doomerists and grifters shouting. Only thing I need is reasonable promise that my data won't be used for training or at least some of it won't. Being able to export conversations in bulk would be helpful.

> Does anyone have any experience with Qwen or any non-Western LLMs?

I use GLM-5.1 for coding hobby project, that going to end up on github anyway. Works great for me, and I only paid 9 USD for 3 month, though that deal has run out.

> my data won't be used for training

Yeah, I don't know. Doubt it.

Re: Qwen3.6-35B-A3B: Agentic coding power, now open to all

#29

Post by dgb23 »

mhitza wrote: post #24
abhikul0 wrote: post #7

I hope the other sizes are coming too(9B for me). Can't fit much context with this on a 36GB mac.

It's a MoE model and the A3B stands for 3 Billion active parameters, like the recent Gemma 4. You can try to offload the experts on CPU with llama.cpp (--cpu-moe) and that should give you quite the extra context space, at a lower token generation speed.

Do I expect the same memory footprint from an N active parameters as from simply N total parameters?

Re: Qwen3.6-35B-A3B: Agentic coding power, now open to all

#30

Post by pdyc »

mhitza wrote: post #24
abhikul0 wrote: post #7

I hope the other sizes are coming too(9B for me). Can't fit much context with this on a 36GB mac.

It's a MoE model and the A3B stands for 3 Billion active parameters, like the recent Gemma 4. You can try to offload the experts on CPU with llama.cpp (--cpu-moe) and that should give you quite the extra context space, at a lower token generation speed.

i dont get it, mac has unified memory how would offloading experts to cpu help?
Post reply on HN