I really hope this doesn't hinder development too much. As Simon says, Qwen3.5 is very impressive.
I've been testing Qwen3.5-35B-A3B over the past couple of days and it's a very impressive model. It's the most capable agentic coding model I've tested at that size by far. I've had it writing Rust and Elixir via the Pi harness and found that it's very capable of handling well defined tasks with minimal steering from me. I tell it to write tests and it writes sane ones ensuring they pass without cheating. It handles the loop of responding to test and compiler errors while pushing towards its goal very well.
I've been testing the same with some rust, and it's has spent a fair bit of time going through an infinite seeming loop before finally unjamming itself. It seems a little more likely to jam up than some other models I've experimented with.
It's also driving itself crazy with deadpool & deadpool-r2d2 that it chose during planning phase.
That said, it does seem to be doing a very good job in general, the code it has created is mostly sane other than this fuss over the database layer, which I suspect I'll have to intervene on. It's certainly doing a better job than other models I'm able to self-host so far.
Some of the early quants had issues with tool calling and looping. So you might want to check that you're running the latest version / recommended settings.
It's the number of active parameters for a Mixture of Experts (misleading name IMO) model.
Qwen3.5-35B-A3B means that the model itself consists of 35 billion floating point numbers - very roughly 35GB of data - which are all loaded into memory at once.
But... on any given pass through the model weights only 3 billion of those parameters are "active" aka have matrix arithmetic applied against them.
This speeds up inference considerably because the computer has to do less operations for each token that is processed. It still needs the full amount of memory though as the 3B active it uses are likely different on every iteration.
In my experience qwen 3 coder next is better. I ran quite a few tests yesterday and it was much better at utilizing tool calls properly and understanding complex code. For its size though 3.5 35B was very impressive. coder next is an 80b model so i think its just a size thing - also for whatever reason coder next is faster on my machine. Only model that is competitive in speed is GLM 4.7 flash
I use the term "harness" for those - or just "coding agent". I think orchestrator is more appropriate for systems that try to coordinate multiple agents running at the same time.
This terminology is still very much undefined though, so my version may not be the winning definition.
In my experience Qwen3.5 is better even at smaller distillations. From what I understand the Qwen3-next series of models was just a test/preview of the architectural changes underpinning Qwen3.5. So Qwen3.5 is a more complete and well trained version of those models.
I would second that Qwen3.5 is exceptionally good. In a calibration, it (35b variant) was running locally with Ada NextGen 24GB to do the same things with easy-llm-cli in comparison with gemini-cli + Gemini 3 Pro, they were at par … really impressive it ran pretty fast …
China is also giving them dump trucks full of cash though. Plus you have to content with the nationalism reason (unfortunately this has died off in America for too many). The idea of building your country is valued for most Chinese I have met. Plus China is incredibly nice to live in, especially if you have lots of money and/or connections. So you can work in China, get paid lots of money, feel like you are doing good. Or In America you can get paid lots of money, and get yelled at by people online because the Government wants to use your model.
China city life is amazingly convenient. Trains and subways are just such an enormous quality of life boost. Add to that the relative cleanliness of having nearly zero homelessness and you’ve got something very compelling.
I will say we are winning in accessibility. China doesn’t have much of a ramp game
They probably have tried, but you have to have more cash than those researchers feel they can get starting their own lab. When you consider the fact that their new startup lab would have the entire nation of China as, in effect, a captive market; you start to see how almost any amount of money would be too little to convince them not to make a run at that new startup. If money is their aim.
I think Alibaba needs to just give these guys a blank check. Let them fill it in themselves. Absent that, I'm pretty sure they'll make their own startup.
I do think it'd be a big loss for the rest of the world though if they close whatever model their startup comes up with.
I use Qwen 3 Coder Next daily on my mac as my main coding agent. It is incredibly capable and its strange how you are painting this picture as if its a fringe use case, there are whole communities that have popped up around running local models.
Can I doubt your claim? I have had such terrible luck with AI coding on <400B models. Not to mention, I imagine your codebase is tiny. Or you are working for some company that isnt keeping track of your productivity.
I am trying super hard to use cheap models, and outside SOTA models, they have been more trouble than they are worth.
Yesterday I test ran Qwen3.5-35B-A3B on my MBP M3 Pro with 36GB via LM Studio and OpenCode. I didn’t have it write code but instead use Rodney (thanks for making it btw!) to take screenshots and write documentation using them. Overall I was pretty impressed at how well it handled the harness and completed the task locally. In the past I would’ve had Haiku do this, but I might switch to doing it locally from now on.
I suppose this shows my laziness because I'm sure you have written extensively about it, but what orchestrator (like opencode) do you use with local models?
I've used opencode and the remote free models they default to aren't awful but definitely not on par with Gemini CLI nor Claude. I'm really interested in trying to find a way to chain multiple local high end consumer Nvidia cards into an alternative to the big labs offering.
I think this is directing coders towards self-sufficiency and that's a good thing. If they don't end up using it for agentic coding, they can use it for running tests, builds, non-agentic voice controlled coding, video creation, running kubernetes, or agent orchestration. So no, it's not evil, even if it doesn't go quite as expected.
Getting a bit of whiplash goin from AI is replacing people, to AI is dead without (these specific) people. Surely we're far enough ahead that AI can take it from here?
I tried the new qwen model in Codex CLI and in Roo Code and I found it to be pretty bad. For instance I told it I wanted a new vite app and it just started writing all the files from scratch (which didn’t work) rather than using the vite CLI tool.
Is there a better agentic coding harness people are using for these models? Based on my experience I can definitely believe the claims that these models are overfit to Evals and not broadly capable.
I've noticed that open weight models tend to hesitate to use tools or commands unless they appeared often in the training or you tell them very explicitly to do so in your AGENTS.md or prompt.
They also struggle at translating very broad requirements to a set of steps that I find acceptable. Planning helps a lot.
Regarding the harness, I have no idea how much they differ but I seem to have more luck with https://pi.dev than OpenCode. I think the minimalism of Pi meshes better with the limited capabilities of open models.
I've been testing Qwen3.5-35B-A3B over the past couple of days and it's a very impressive model. It's the most capable agentic coding model I've tested at that size by far. I've had it writing Rust and Elixir via the Pi harness and found that it's very capable of handling well defined tasks with minimal steering from me. I tell it to write tests and it writes sane ones ensuring they pass without cheating. It handles the loop of responding to test and compiler errors while pushing towards its goal very well.
It's also driving itself crazy with deadpool & deadpool-r2d2 that it chose during planning phase.
That said, it does seem to be doing a very good job in general, the code it has created is mostly sane other than this fuss over the database layer, which I suspect I'll have to intervene on. It's certainly doing a better job than other models I'm able to self-host so far.
Qwen3.5-35B-A3B means that the model itself consists of 35 billion floating point numbers - very roughly 35GB of data - which are all loaded into memory at once.
But... on any given pass through the model weights only 3 billion of those parameters are "active" aka have matrix arithmetic applied against them.
This speeds up inference considerably because the computer has to do less operations for each token that is processed. It still needs the full amount of memory though as the 3B active it uses are likely different on every iteration.
This terminology is still very much undefined though, so my version may not be the winning definition.
I will say we are winning in accessibility. China doesn’t have much of a ramp game
I think Alibaba needs to just give these guys a blank check. Let them fill it in themselves. Absent that, I'm pretty sure they'll make their own startup.
I do think it'd be a big loss for the rest of the world though if they close whatever model their startup comes up with.
That's very likely to happen once the gap with OpenAI/Anthropic has been closed and they managed to pop the bubble.
Isnt it interesting that you never see someone say "I used this on my Mac and it was useful"
Instead we get "you could put this on your Mac" or "I tried it, and it worked but it was too slow"
I feel like these people are performing an evil when they are making suggestions that cause a waste of money.
I am trying super hard to use cheap models, and outside SOTA models, they have been more trouble than they are worth.
Maybe Qwen3.5-35B-A3B is that model? This comment reports good results: https://news.ycombinator.com/item?id=47249343#47249782
I need to put that through its paces.
So far none of them have be useful enough at first glance with a local model for me to stick with them and dig in further.
To be honest, it's sort of what I expected governments to be funding right now, but I suppose Chinese companies are a close second.
Wild times!
Is there a better agentic coding harness people are using for these models? Based on my experience I can definitely believe the claims that these models are overfit to Evals and not broadly capable.
They also struggle at translating very broad requirements to a set of steps that I find acceptable. Planning helps a lot.
Regarding the harness, I have no idea how much they differ but I seem to have more luck with https://pi.dev than OpenCode. I think the minimalism of Pi meshes better with the limited capabilities of open models.
the qwen is dead, long live the qwen.
https://news.ycombinator.com/item?id=47246746