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.

+1 to this, anecdotally I’ve found in my own evaluations that if your system prompt doesn’t explicitly declare how to invoke a tool and e.g. describe what each tool does, most models I’ve tried fail to call tools or will try to call them but not necessarily use the right format. With the right prompt meanwhile, even weak models shoot up in eval accuracy.

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.

what's your take between Qwen3.5-35B-A3B and Qwen3-Coder-Next?

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

What do you use as the orchestrator? By this I mean opencode, or the like. Is that the right term?

Another vote in favour of "harness".

I'm aligning on Agent for the combination of harness + model + context history (so after you fork an agent you now have two distinct agents)

And orchestrator means the system to run multiple agents together.

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.

What is the meaning of 'A3B'?

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.

It will benefit from a full amount of memory for sure, but AIUI if you use system memory and mmap for your experts you can execute the model with only enough memory for the active parameters, it's just unbearably slow since it has to swap in new experts for every token. So the more memory you have in excess to that, the more inactive but often-used experts can be kept in RAM for better performance.

The ability to stream weights from disk has nothing to do with MoE or not. You can always do this. It will be unusable either way.

Agreed but for a dense model you'd have to stream the whole model for every token, whereas with MoE there's at least the possibility that some experts may be "cold" for any given request and not be streamed in or cached. This will probably become more likely as models get even sparser. (The "it's unusable" judgmemt is correct if you're considering close-to-minimum reauirements, but for just getting a model to fit, caching "almost all of it" in RAM may be an excellent choice.)

Unlike offloading weights from VRAM to system RAM, I just can't see a situation where you would want to offload to an SSD. The difference is just too large, and any model so large you can't run it in system RAM, is going to be so large it is probably unusable except in VRAM.

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...

Have you tried the '--jinja' flag in llama-server?

I’ve been running it via llama-server with no issues. Running the latest Bartowski 6-bit quant

Bartowski? Like Chuck Bartowski from the TV show?

Different one. Bartowski is a minor celebrity in the local LLM world, together with Unsloth.

What's the selling point of these quants vs the Unsloth ones?

What hardware do you have it running on? Do you feel you could replace the frontier models with it for everyday coding? Would/will you?

I'm getting ~30 tok/s on the A3B model with my 3070 Ti and 32k context.

Do you feel you could replace the frontier models with it for everyday coding? Would/will you?

Probably not yet, but it's really good at composing shell commands. For scripting or one-liner generation, the A3B is really good. The web development skills are markedly better than Qwen's prior models in this parameter range, too.

60 to 70 on a 5080, but only tinkering for now. The smaller models seem exceptionally good for what they are, and some can even do OCR reliably.

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.

Turn down the temperature and you’ll see less “simpler” short cuts.

For the uninitiated: Interestingly, it is not advisable to take this to the extreme and set temperature to 0.

That would seem logical, as the results are then completely deterministic, but it turns out that a suboptimal token may result in a better answer in the long run. Also, allowing for a little bit of noise gives the model room to talk itself out of a suboptimal path.

I like to think of this like tempering the output space. With a temperature of zero, there is only one possible output and it may be completely wrong. With even a low temperature, you drastically increase the chances that the output space contains a correct answer, through containing multiple responses rather than only one.

I wonder if determinism will be less harmful to diffusion models because they perform multiple iterations over the response rather than having only a single shot at each position that lacks lookahead. I'm looking forward to finding out and have been playing with a diffusion model locally for a few days.

That sounds awfully similar to what Opus 4.6 does on my tasks sometimes.

Blah blah blah (second guesses its own reasoning half a dozen times then goes). Actually, it would be a simpler to just ...

Specifically on Antigravity, I've noticed it doing that trying to "save time" to stay within some artificial deadline.

It might have something to do with the system messages and the reinforcement/realignment messages that are interwoven into the context (but never displayed to end-users) to keep the agents on task.

Yeah that happened to me with Claude code opus 4.6 1M for the first time today. I had to check the model hadn’t changed. It was weird. I was imagining that maybe anthropic have a way of deciding how much resource a user actually gets and they had downgraded me suddenly or something.

Claude Code recently downgraded the default thinking level to “medium”, so it’s worth checking your settings.

Thank you thank you you give me hope :)

But how do you see the current thinking level and how do you change it? I’ve been clicking around and searching and adding “effortLevel”:”high” to .claude/settings.json but no idea if this actually has any effect etc.

As per Anthropic support (for Mac and Linux respectively) -

  $ echo 'export ANTHROPIC_EFFORT="high"' >> ~/.zshrc source ~/.zshrc
  $ echo 'export ANTHROPIC_EFFORT="high"' >> ~/.bashrc source ~/.bashrc

I prefer settings.json (VSCode) -

  "claudeCode.environmentVariables": [
    { "name": "ANTHROPIC_MODEL", "value": "claude-opus-4-6" },
    { "name": "CLAUDE_CODE_EFFORT_LEVEL", "value": "high" }
  ], ...

recently being within the past 24 hours lol: https://github.com/anthropics/claude-code/releases/tag/v2.1....

As someone that started using Co-work, I feel like I am going insane with the frequency that I have to keep telling it to stay on task.

If you ask it to do something laborious like review a bunch of websites for specific content it will constantly give up, providing you information on how you can continue the process yourself to save time. Its maddening.

In my experience all of the models do that. It's one of the most infuriating things about using them, especially when I spend hours putting together a massive spec/implementation plan and then have to sit there babysitting it going "are you sure phase 1 is done?" and "continue to phase 2"

I tend to work on things where there is a massive amount of code to write but once the architecture is laid down, it's just mechanical work, so this behavior is particularly frustrating.

I hope you will excuse my ignorance on this subject, so as a learning question for me: is it possible to add what you put there as an absolute condition, that all available functions and data are present as an overarching mandate, and it’s simply plug and chug?

Recently it seems that even if you add those conditions the LLMs will tend to ignore them. So you have to repeatedly prompt them. Sometimes string or emphatic language will help them keep it “in mind”.

That’s pretty funny when compared with the rhetoric like “AI doesn’t get tired like humans.” No, it doesn’t, but it roleplays like it does. I guess there is too much reference to human concerns like fatigue and saving effort in the training.

This is what happens when a bunch of billionaires convince people autocomplete is AI.

Don't get me wrong, it's very good autocomplete and if you run it in a loop with good tooling around it, you can get interesting, even useful results. But by its nature it is still autocomplete and it always just predicts text. Specifically, text which is usually about humans and/or by humans.

You are not wrong, but after having started working with LLMs, I have this feeling that many humans are simply autocomplete engines too. So LLMs might be actually close to AGI, if you define "general" as "more than 50% of the population".

to decide halfway through following my detailed instructions that it would be "simpler" to just... not do what I asked

That's likely coming from the 3:1 ratio of linear to quadratic attention usage. The latest DeepSeek also suffers from it which the original R1 never exhibited.

I've had even better results using the dense 27B model -- less looping and churning on problems

Which dense model are you referring to? The dense model isn’t finetuned for code instruction according to the model card.

https://huggingface.co/Qwen/Qwen3.5-27B

I wasn't aware of that, which page mentions that?

What hardware are you running this on?

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?

Wild times!

Anthropic has one nine of uptime right now. One.

https://status.claude.com/

If AI could effectively replace people, you wouldn’t need CEOs to keep trying to convince people.

Not sure what the uptime is meant to signal. People have quite low uptime as well…

That's 99% is two nines?

It was 98.xx this morning when I posted.

This is pretty pedantic, but I think it's usually rounded. 1=90%, 2=99%, 3=99.9%. I'd say 98% is "not even two nines" but not "one nine".

9% uptime?

Everything on that page has two nines, so not sure what you're trying to say here.

"product market fit is when people are ripping the product out of your hands and everything is breaking constantly" - seems bullish to me

I wonder how a US lab hasn't dumped truckloads of cash into various laps to ensure these researchers have a place at their lab

Indeed; or, Europe badly needs a competitive model to hedge against US political nonsense.

Offering „You are welcome“ relocation package to Anthropic might be a good idea.

Given how American govt. has treated Anthropic, I think you might be right. EU truly has a remarkable opportunity to make Anthropic/Claude European.

This US administration (or any admin) would almost certainly impose export controls on US AI technology before it would allow one of the frontier model providers to be acquired/relocate outside the US. It did the same thing when ASML wanted to acquire Cymer (California company that provides the EUV light source technology). The acquisition was only allowed under strict technology sharing/export agreements with the Dutch government.

Europe really just needs to rally behind Mistral. That's where they should dump their cash.

Having one „champion“ is flawed European approach. We need local competition and headhunting to make it fly.

Hard to compete in an environment that’s anti-996 and the pay is so much less.

Do you consider high working hours to be a benefit akin to higher pay? I think fewer hours and less money is a fair deal for employees.

There is no capital in the EU

Anthropic has gone out of their way to make a point about how much they love and admire the US state and its defense sector. Only drawing the line at a very far point and even when they drew the line it was with a big thing about how they believe in the American defense sector blah blah blah.

In any case, there's no way Anthropic's investors in Silicon Valley would countenance such a move.

Also, I'm biased the logical place is Canada, not Europe. Much of the fundamental/foundational research on LLMs, and a large part of the talent, came from universities in Canada anyways.

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 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.

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

All very true.

I wonder if you max out your options in China. It seems the Party is suspicious of ambition and high profile winners. I'm sure you can live comfortably, but there's a ceiling.

what is the issue with having a ceiling?

Star athletes really hate being told they can't score more than 10 goals in a season because it's unfair to the other weaker players. The players will either leave to go play somewhere else, or they become weaker players themselves.

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.

Isn't it just straight-up illegal in China to refuse the government from using your model? USA isn't perfect, but at least it has active discourse.

I would imagine if it isn't illegal its a very bad idea not to. But regardless, I would bet large amounts of money that you would never get any flack for doing anything for the government. If I went on X, Threads, Bluesky, TikTok and said "Hey I am a software engineer selling awesome new technology to the government and military!" I am going to get Americans attacking me for supporting Trump / ICE / FBI whatever the current issue of the day is. If I did the same on Douyin or Weibo the response would be able making China strong, and there would be no criticism of that choice.

Sure, but the difference is that while the Chinese state is measurably awful on all sorts of human rights things within their own borders... they're not currently dropping bombs on foreign cities, starving a neighbour of critical petroleum shipments, or heavily funding an ally to slowly exterminate a population.

What point are you trying to make here? Are government abuses somehow inherently better or worse depending on where they happen?

Do you imagine an invasion of Taiwan won't involve dropping bombs?

I feel like we should be able to agree that providing authoritarian regimes with high tech tools is immoral in the general case.

My point is as a non-American I feel no allegiance to either state, and current events don't make me sympathetic to the geo-political aims of the USA. So I don't see a strong moral case for this tech being an especial purvey of either party.

If you'd asked me two years ago my answer might have been different.

And to the original point, yeah, I would feel entirely justified in the critique of engineers in providing tools to the US defense apparatus at this point.

At least the Chinese shops are giving their weights away for free, and not demanding that any government ban the rest.

Why are they giving them away for free?

At least it has been decades since China Gov bombed innocent people in other countries. A peaceful and responsible government.

A peaceful and responsible government.

People in Hong Kong died. Over 10,000 were arrested and many are still in prison. The rest are permanently disgraced in their social-credit society.

Again, USA is not perfect, but let's not dream up some fantasy about the CCP.

This "social credit" thing is dead in China.

As an American, I have no fear of calling the US President a pedo or saying Fuck the Police on my Twitter. Not the case in China. It's horrifying.

https://reclaimthenet.org/china-man-chair-interrogation-soci...

Oh, China absolutely does not tolerate _public_ dissent very much including highly visible social media posts. Everybody there knows that.

But this:

According to the social credit system, Chinese citizens are punishable if they indulge in buying too many video games, buying too much junk food, having a friend online who has a low credit score, visiting unauthorized websites, posting “fake news” online, and more.

...is just pure bullshit. There were _ideas_ about including these kinds of stuff into the score, but they have never been implemented. At this point, the social credit score is only used to find people who dodge court decisions.

I have no fear of calling the US President a pedo or saying Fuck the Police on my Twitter.

Does that matter? In China people don't judge the state of their civilization by how easily you can insult the police but whether you need to be afraid to meet them on the street. "I can insult my pedophile president" (who doesn't care if you do) isn't exactly a flex.

It does tell us something though that the evaluation of American life now consists of parasocial interactions with the president on social media. I'm starting to belief Bruno Maçães, ex Portuguese secretary of state, was prescient with his diagnosis that American material society has rotted to the point where life is now entirely defined by virtual interactions. That's the difference between China and the US today.

The president's a pedophile, a criminal, undeterred by democracy, economy or social disorder but you can freely yell into the void. Have you considered that in the US one can freely say all these things precisely because that's irrelevant?

People in Hong Kong died

Do you have a legit source for this? When I search for information, I only found this case, “Luo Changqing, a 70-year-old Hong Kong cleaner, died from head injuries sustained after he was hit by a brick thrown by a Hong Kong protester during a violent confrontation between two groups in Sheung Shui, Hong Kong on 13 November 2019.”

None of the other legit sources claim the police killed any of the rioters.

Constant military drills around Taiwan isn't peaceful or responsible.

China is bullying lots of countries in the SCS (ramming Philippine coast guard ships, building military installations in the SCS, ...). Not peaceful or responsible.

If China was serious about a military solution for Taiwan they would be invading right now while the US is unloading into the desert.

China is incredibly nice to live in

I'm sure it's a very nice place to live if you're content to just stay quiet in society and never put a political sign in your yard or even just talk about the wrong thing with your friend in a WeChat.

try to protest in america and see how that works out for you long-term. or say protest against genocide in gaza at an uni or generally in public…

Sigh. Let's not invent things? You can protest anything in the US just fine, with generally no consequences. Heck, our local _high_ _school_ students go out and protest everything to weasel out of classes.

Trump admin did put people in prison and then deported them, for doing nothing more than protesting.

Not as bad as China sure, but not as good as other civilized nations.

Let's just clarify that visitors don't have the same rights as citizens. Whether or not you agree with the current administration's policies hopefully we can agree that it is entirely reasonable for them to deport foreign political dissidents more or less at their discretion.

If you want to put this to the test try crossing the Canadian border and when they ask you the purpose of your visit respond that it's to attend a protest.

Let's just clarify that visitors don't have the same rights as citizens.

Yunseo Chung was not a visitor. She came to the United States from South Korea at age 7. She was arrested last year for peacefully protesting. Charges against her were dropped but the govt. canceled her green card.

The govt. has been trying to deport her since then, but the courts keep blocking it.

https://humanrightsfirst.org/yunseo-chung-v-trump-administra...

While the legality of these actions are being debated in courts, I think most of us can agree that this is reprehensible behavior on part of the Trump admin.

But the constitution is not worded as if they don't have the same fundamental rights. Even in other countries, it is the same; this is done to prevent slavery and unjust incarcelation. So visitors have the same fundamental rights to free speech, fair trial, etc. The US has also agreed to international conventions. But the current administration seems to not care

Trump admin did put people in prison and then deported them, for doing nothing more than protesting.

Link? I’m guessing we’re going to see that this definition of “protesting” involves being aggressive and directly in the face of law enforcement officers, not merely holding a sign at a distance.

this is funny if you are being sarcastic

This is an exaggeration. Nobody in China cares about what you speak with each other privately, and people talk about stupid policies all the time. The government cares about _public_ actions.

In practical terms, if you're not kind of person who would want to run for an office in the US, China is incredibly comfortable. Cities are safe, with barely any violent crime. Public drug use is nonexistent. And with the US-level AI researcher income, you'd be in the top 0.1% earners.

nobody in China cares about what you speak with each other privately, and people talk about stupid policies all the time. The government cares about _public_ actions.

https://news.ycombinator.com/item?id=47252833

My comment and the linked video says otherwise. The guy was in a private group chat and said some nasty things about the police for confiscating his motorcycle. Now he's arrested and in the Tiger Chair.

How are we explaining this?

Chinese people are very racist towards non-Chinese. It might seem like a happy utopia, but if you aren't Chinese, then you may not really enjoy your time there. It may not be quite as bad as being black in rural US south, but being black (or anything non-Chinese) in China is still not going to be a good time.

Wild to call 1.42 billion people racist despite having met very few of them.

Racism in even the worse parts of America doesn't even begin to touch the racism present in monocultural/monoracial countries.

Have you experienced racism? In Japan atleast, it was evenly applied. That company won't rent to foreigners but this one will. That company won't hire foreigners but this one will. Police will bother you if you ride a bike, but they will be polite while they waste 10 minutes of your time asking for your gaijin card for biking while foreign.

In the US people try to hide it and are far more sinister about it, since there are a lot of laws against obvious racism. The cops are also happy in the US to just kill you.

The racism in the US comes out of hate where as what I experienced abroad was more, we don't think you'll fit in and follow the rules and you have to constantly prove that you can.

I didn't spend too much time in China so maybe it is a racist hell hole.

But my experience in Japan was that white immigrants were way more inclined to make a huge deal about the lighter racism they experienced because they had never been somewhere where their skin color was a disadvantage.

This is a weird argument. Japanese racism is fine because the Japanese are polite and apply it evenly?

There's a very big difference between xenophobia and racism. Racism is much worse.

What do you mean by racist? I'm a white/hispanic American and spent 3 months in China and didn't really notice anything problematic towards me.

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.

Yes. Yes, so true. And the phd types building these models are probably even scared in China that ICE will fly there to deport them.

This thread is about bringing these people to the US.

This thread is largely pointless political back-and-forth were predictably the comments with a more positive opinion on current US immigration enforcement will be flagged.

To get back to the original point, personally I doubt sentiment on US immigration enforcement would be so significant a deterrent for Chinese talent, who may not share the political views of the American left for whom this is a big concern.

Yeah, the Hyundai factory fiasco kind of dashed the idea that the enforcement would spare people working in favored industries setting up in the US.

The Hyundai factory "enforcement" wasn't even legal. Those workers were here to train US workers and the Hyundai employees had proper visas for this.

https://apnews.com/article/immigration-raid-hyundai-korea-ic...

https://www.koreatimes.co.kr/foreignaffairs/20251112/hundred...

https://www.pbs.org/newshour/nation/attorney-says-detained-k...

The regime is powered by racism and doesn't think through things.

"Papers, please." comes to the US of A.

Every administration since the foundation of ICE has removed illegal immigrants and funded ICE and immigration policy/border operations [1].

[1] Removals by president: https://www.migrationpolicy.org/article/biden-deportation-re...

You have to show ID to pick up a prescription or open a bank account. You have to show ID for routine traffic stops. This is such a juvenile, tired argument.

“You show ID at the bank” is a classic, juvenile and tired argument because it swaps in a voluntary transaction for state coercion.

The concern isn’t IDs exist—it’s who’s demanding them, in what context, and what happens if you can’t comply on the spot.

You absolutely do not have to show ID to pick up every prescription; just some, which is also dependent on state law, federal law, and pharmacy.

But also, I don't care if it's a tired argument--this isn't about how things are, it's about how we want them to be. I don't want to live in a state action-coerced society.

In Sweden you have to use your Bank ID to take the buss. Meaning the bank has the same security as entering the bus. If you get robbed in Sweden they take your bank ID, use it to take a loan and then transfers all your money to a foreign bank. Because we got rid of manual cash.

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.

Where do researchers go instead of USA then? Genuinely curious

Were they kneecapped by Anthropic blocking their distillation attempts?

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.

The "distillation attacks" are mostly using Claude as LLM-as-a-judge. They are not training on the reasoning chains in a SFT fashion.

So they're paying expensive input tokens to extract at best a tiny amount of information ("judgment") per request? That's even less like "distillation" than the other claim of them trying to figure out reasoning by asking the model to think step by step.

My conspiracy theory hat is that somehow investors with a stake in openai as well is sabotaging, like they did when kicking emad out of stabilityai

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.

source?

Does anyone know when the small Qwen 3.5 models are going to be on OpenRouter?

they're already there ?? https://openrouter.ai/qwen/qwen3.5-27b

Like 4B, 2B, 9B. Supposedly they are surprisingly smart.

Yep. The 9B has excellent image recognition. I showed it a PCB photo and it correctly identified all components and the board type from part numbers and shape. OCR quality was solid. Tool calling with opencode worked without issues, but general coding ability is still far from sonnet-tier. Asked it to add a feature to an existing react app, it couldn't produce an error-free build and fell into a delete-redo loop. Even when I fixed the errors, the UI looked really bad. A more explicit prompt probably would have helped. Opus one-shotted it, same prompt, the component looked exactly as expected.

But I'll be running this locally for note summarization, code review, and OCR. Very coherent for its size.

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.

Wasted tokens are preferred for local models, I need the GPU mainframe in my bedroom to heat it as I live in a third world country with unreliable heating (Switzerland).

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...)

How do those companies make money? Qwen, GLM, Kimi, etc all released for free. I have no experience in the field, but from reading HN alone my impression was training is exceptionally costly and inference can be barely made profitable. How/why do they fund ongoing development of those models? I'd understand if they release some of their less capable models for street cred, but they release all their work for free.

Ostensibly, a mix of VC funding and that they host an endpoint that lets them run the big (200+GB) models on their infrastructure rather than having to build machines with hundreds of gigs of llm-dedicated memory.

But on inference they have to compete with other inference provider that just has a homepage, a bunch of GPUs running vllm and none of the training cost. Their only real advantage are the performance optimizations that they might have implemented in their inference clusters and not made public

Chinese companies don't always operate on purely capitalistic principles, there is sometimes government direction in the background.

For China, the country, it's a good thing if American AI companies have to scramble to compete with Chinese open models. It might not be massively profitable for the companies producing said models, but that's only a part of the equation

China seems to combine the best points of capitalism (many companies taking many shots on goal, instead of the eastern bloc way of one centrally-mandated solution that either works or not) with the best points of communism (state-sponsored industries that don't have to generate a profit, for the glory and benefit of the state).