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.
Have frontier lab do the plan which is the most time consuming part anyways and then local llm do the implementation.
Frontier model can orchestrate your tickets, write a plan for them and dispatch local llm agents to implement at about 180 tokens/s, vllm can probably ,manage something like 25 concurrent sessions on RTX 6000
Do it all in a worktrees and then have frontier model do the review and merge.
I am just a retired hobbyist but that's my approach, I run everything through gitea issues, each issue gets launched by orchestrator in a new tmux window and two main agents (implementer and reviewer get their own panes so I can see what's going on). I think claude code now has this aspect also somewhat streamlined but I have seen no need to change up my approach yet since I am just a retired hobbyist tinkering on my personal projects. Also right now I just use claude code subagents but have been thinking of trying to replace them with some of these Qwen 3.5 models because they do seem cpable and I have the hardware to run them.
In my experience Qwen3.5/Qwen3-Coder-Next perform best in their own harness, Qwen-Code. You can also crib the system prompt and tool definitions from there though. Though caveat, despite the Qwen models being the state of the art for local models they are like a year behind anything you can pay for commercially so asking for it to build a new app from scratch might be a bit much.
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.
Are you running it locally with llama.cpp? If so, is it working without any tweaking of the chat template? The tool calls fail for me when using the default chat template, however it seems to work a whole lot better with this: https://huggingface.co/Qwen/Qwen3.5-35B-A3B/discussions/9#69...
I've been playing with 3.5:122b on a GH200 the past few days for rust/react/ts, and while it's clearly sub-Sonnet, with tight descriptions it can get small-medium tasks done OK - as well as Sonnet if the scope is small.
The main quirk I've found is that it has a tendency to decide halfway through following my detailed instructions that it would be "simpler" to just... not do what I asked, and I find it has stripped all the preliminary support infrastructure for the new feature out of the code.
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?
ICE has been detaining Chinese people in my area (and going door to door in at least one neighborhood where a lot of Chinese and Indians live). I was hearing about this just last week as word spread amongst the Chinese community here (Ohio) to make sure you have some legal documentation beyond just your driver's license on you at all times for protection. People will hear about this through the grapevine and it has a massive (and rightly so) chilling effect. US labs can try but with US government behaving like it is I don't think they will have much luck.
*edit: not that it matters, but since MAGA can't help but assume, these are all US citizens and green card holders that I am referring to.
There are 2 groups of new Chinese immigrants in the US, they are quite different:
1.Those who arrived through legal channels (most studied at U.S. universities and remained on H1B visas, with a smaller number through EB5 or other visa categories) and eventualy got green card.
2.Undocumented immigrants, which include several sub-groups/waves. In the 1990s, most came from just a couple provinces, Fujian and southern Zhejiang. After COVID, they were from different parts of China and entered through the southern border.
The contributors to AI development belong to the 1st group. They are spread across the country but a large number work in high-tech companies in Northern California.
The 2nd group was intially concentrated in New York and Southern California (Los Angeles area). Later they have expanded into nearby regions. They provide labor for Chinese-owned small businesses such as restaurants, grocery stores, and hotels.
There is an industry created largely by Chinese political dissidents helping Group 2 through asylum applications using fake materials and exploiting common Western beliefs or narratives about China like human rights concerns. For example, Alysa Liu’s father is an asylum lawyer.
ICE enforcement efforts would likely focus more on Group 2 if they are knowledgable. Ohio should not be a high-priority area. I could be wrong due to changes over time. One indicator you can observe: Are there many Chinese-owned small businesses in your area?
Well, the problem aren't just the NSF funding cuts. Everyone else is already dumping truckloads of cash. There's also the public health situation (who wants measles or polio?), the risk of retaliatory attacks from the countries we're at war with, etc. You could write paragraphs about why the US is less attractive to researchers.
When I was a deep learning PhD in the first Trump administration, US universities were already very deeply affected by the Muslim ban, and so a lot of talent ended up in other countries.
Sibling commentators are rightfully pointing out that foreigners, especially those who would not be recognized as white, face an onerous and risky customs process with long-term and increasing risks of deportation. When you see a headline like the NIST labs abruptly restricting foreign scientists, _everything_ else feels uncertain. Even if someone doesn't believe they're personally at risk for deportation, they're still seeing everything else.
And then it all boils down to a reputational thing. The era where we were the top choice for research is in the past. If you start a PhD in the US on your resume during this era, you might be anticipating how you'll answe the question of why you weren't good enough to get accepted somewhere better.
They already kind of do, but I think anyone who was into US money has already left for it, and the money China is throwing at the problem is pretty good also. You can also have a lot more influence in a Chinese company without having to adopt a weird new American corporate culture.
What Anthropic was complaining about is training on mass-elicited chat logs. It is very much a ToS violation (you aren't allowed to exploit the service for the purpose of building a competitor) so the complaint is well-founded but (1) it's not "distillation" properly understood; it can only feasibly extract the same kind of narrow knowledge you'd read out from chat logs, perhaps including primitive "let's think step by step" output (which are not true fine-tuned reasoning tokens); because you have no access to the actual weights; and (2) it's something Western AI firms are very much believed to do to one another and to Chinese models all the time anyway. Hence the brouhaha about Western models claiming to be DeepSeek when they answer in Chinese.
I managed to get qwen2.5-coder:14B working under ollama on an Nvidia 2080 Ti with 11GB of VRAM, using ollama cli, outputting what looks like 200 words-per-minute to my eye
It has been useful for education ("What does this Elixir code do? <Paste file> ..... <general explanation> "then What this line mean?")
as well as getting a few basic tests written when I'm unfamiliar with the syntax. ("In Elixir Phoenix, given <subject under test, paste entire module file> and <test helper module, paste entire file> and <existing tests, pasted in, used both for context and as examples> , what is one additional test you would write?")
This is useful in that I get a single test I can review, run, paste in, and I'm not using any quota. Generally I have to fix it, but that's just a matter of reading the actual test and throwing the test failure output to the LLM to propose a fix. Some human judgement is required but once I got going adding a test took 10 minutes despite being relatively unfamiliar with Elixir Phoenix .
It's a nice loop, I'm in the loop, and I'm learning Elixir and contributing a useful feature that has tests.
More likely some high ranking party member's nepobaby from Gemini sniffed success with Qwen and the original folks just walked away as their reward disappeared.
apples v.s. oranges. The later is true, Emad did get sabotaged (for not being able to raise money in time, about 8-month before he's leaving). Junyang didn't have that long arc of incidents.
Interesting reading this. It reminds me of my time in cryptocurrency sector. I suspected that some team members were paid by Ethereum folks to sabotage our project. Why do I suspect Ethereum? Because our project founders ended up switching to the Ethereum ecosystem and ignored/suppressed better solutions from their own ecosystem. I think there's something about tech hype which attracts these kinds of people who like to play dirty.
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 …
There has been tension between Qwen's research team and Alibaba's product team, say the Qwen App. And recently, Alibaba tried to impose DAU as a KPI. It's understandable that a company like Alibaba would force a change of product strategy for any number of reasons. What puzzled me is why they would push out the key members of their research team. Didn't the industry have a shortage of model researchers and builders?
I wonder if an american company poached one/all of them. They've been pretty much bleeding edge of open models and would not surprise me if Amazon or Google snatched them up
One thing I’ve noticed with local models is that people tolerate a lot more trial and error behavior. When a hosted model wastes tokens it feels expensive, but when a local model loops a bit it just feels like it’s “thinking.”
If models like Qwen can get good enough for coding tasks locally, the real shift might be economic rather than purely capability.
It sounds like the lead was demoted to attract new talent, quit as a result, and the rest of the team also resigned to force management to change their minds.
If so, I'm happy that the team held together, and I hope that endogenous tech leads get to control their own career and tech destiny after hard work leads to great products. (It's almost as inspiring as tank man, and the tank commanders who tried to avoid harming him...)
(ducking the downvote for challenging the primacy of equity...)
As a mathematician, lately I experimented a lot with Qwen, to produce as good as possible professional summaries and relations between articles, and in one case even a verification of misattributions claims which was used in an arXiv article.
56 comments
[ 6.1 ms ] story [ 86.1 ms ] threadthe qwen is dead, long live the qwen.
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.
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.
https://news.ycombinator.com/item?id=47246746
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.
The main quirk I've found is that it has a tendency to decide halfway through following my detailed instructions that it would be "simpler" to just... not do what I asked, and I find it has stripped all the preliminary support infrastructure for the new feature out of the code.
Wild times!
https://status.claude.com/
If AI could effectively replace people, you wouldn’t need CEOs to keep trying to convince people.
*edit: not that it matters, but since MAGA can't help but assume, these are all US citizens and green card holders that I am referring to.
1.Those who arrived through legal channels (most studied at U.S. universities and remained on H1B visas, with a smaller number through EB5 or other visa categories) and eventualy got green card.
2.Undocumented immigrants, which include several sub-groups/waves. In the 1990s, most came from just a couple provinces, Fujian and southern Zhejiang. After COVID, they were from different parts of China and entered through the southern border.
The contributors to AI development belong to the 1st group. They are spread across the country but a large number work in high-tech companies in Northern California.
The 2nd group was intially concentrated in New York and Southern California (Los Angeles area). Later they have expanded into nearby regions. They provide labor for Chinese-owned small businesses such as restaurants, grocery stores, and hotels.
There is an industry created largely by Chinese political dissidents helping Group 2 through asylum applications using fake materials and exploiting common Western beliefs or narratives about China like human rights concerns. For example, Alysa Liu’s father is an asylum lawyer.
ICE enforcement efforts would likely focus more on Group 2 if they are knowledgable. Ohio should not be a high-priority area. I could be wrong due to changes over time. One indicator you can observe: Are there many Chinese-owned small businesses in your area?
When I was a deep learning PhD in the first Trump administration, US universities were already very deeply affected by the Muslim ban, and so a lot of talent ended up in other countries.
Sibling commentators are rightfully pointing out that foreigners, especially those who would not be recognized as white, face an onerous and risky customs process with long-term and increasing risks of deportation. When you see a headline like the NIST labs abruptly restricting foreign scientists, _everything_ else feels uncertain. Even if someone doesn't believe they're personally at risk for deportation, they're still seeing everything else.
And then it all boils down to a reputational thing. The era where we were the top choice for research is in the past. If you start a PhD in the US on your resume during this era, you might be anticipating how you'll answe the question of why you weren't good enough to get accepted somewhere better.
Besides you can live a comfortable life in PRC nowadays or live in a racist America.
It has been useful for education ("What does this Elixir code do? <Paste file> ..... <general explanation> "then What this line mean?")
as well as getting a few basic tests written when I'm unfamiliar with the syntax. ("In Elixir Phoenix, given <subject under test, paste entire module file> and <test helper module, paste entire file> and <existing tests, pasted in, used both for context and as examples> , what is one additional test you would write?")
This is useful in that I get a single test I can review, run, paste in, and I'm not using any quota. Generally I have to fix it, but that's just a matter of reading the actual test and throwing the test failure output to the LLM to propose a fix. Some human judgement is required but once I got going adding a test took 10 minutes despite being relatively unfamiliar with Elixir Phoenix .
It's a nice loop, I'm in the loop, and I'm learning Elixir and contributing a useful feature that has tests.
If models like Qwen can get good enough for coding tasks locally, the real shift might be economic rather than purely capability.
If so, I'm happy that the team held together, and I hope that endogenous tech leads get to control their own career and tech destiny after hard work leads to great products. (It's almost as inspiring as tank man, and the tank commanders who tried to avoid harming him...)
(ducking the downvote for challenging the primacy of equity...)
All is collected in https://imar.ro/~mbuliga/ai-talks.html