China just released an open source AI model that matches the best closed models from OpenAI and Anthropic. Gavin Baker explained exactly how they did it and the answer should concern every American AI lab. The model is called GLM 5.2. It was built by Z. AI. You get... 744 billion parameters, 1 million token context window and its MIT license, meaning anyone can download it, fork it, build a company on it, with no restrictions and no Dario. It scored 51 points on the artificial analysis intelligence index. The highest score any open weight model has ever achieved. It beat GPT 5.5 on the frontier software engineering benchmark. It trails Claude Opus 4.8 by less than one percentage point. And it costs 85% less to run than GPT 5.5 for comparable performance. Gavin Baker said on the All-In podcast that this model has challenged some of his beliefs. Then he explained how China built it. The method is called distillation. Just think of tens of thousands of phones and computers running simultaneously, all hitting the frontier model APIs through masked accounts, asking specific questions, and harvesting what happens inside the model when it answers. Every reasoning step, every token. The entire thinking process gets recorded and fed back into the Chinese model during training. It is a cheat sheet. It is the answer key to the exam. And here is the part that should worry everyone. Sacks said it plainly. China was already nine months behind American models. But now that GLM 5.2 is good enough to run its own reinforcement learning, it can improve itself without needing to distill from American models anymore. The cheat sheet let them get close enough to start writing their own answers. Sacks said we are six months behind on the model and 24 months behind on silicon and they are only a few months behind in total. The Z. AI founder told Elon Musk directly that open weight fable-level capability will be here before Q1 2027. Every restriction Anthropic lobbied for, every self-imposed safety guardrail, every month of delay in releasing American frontier models accelerated this. The Chinese labs were not under those restrictions. They were not going to wait. The composable model future Gavin described, where every enterprise runs a frontier model alongside their own fine-tuned open weight model, is coming regardless of what American labs do next. The question is just whether the open weight half of that stack is American or Chinese. Right now it is Chinese. WATCH THE FULL PODCAST ON The All-In Podcastshow more

If they can build better model on reasoning tracks - then there is more than just distillation.

I was explaining the same thing to my friend and he agreed with me.

100%

Open-source AI is becoming the biggest force in the global AI race.

mit license with no dario is doing a lot of heavy lifting here

Haha, we don’t need Dario. Let the world be Dario-free.

The US isn't quite ready yet to come to terms with the fact that China has really stepped on the gas and has all the talent it needs to do so.

Anthropic doesn’t provide a distillation endpoint via their API so this whole narrative and accusation by people including Anthropic is fundamentally flawed and the fact that the people with the loudest voices aren’t picking it up should concern everyone that we’re entering a new level of society where basic thought and fact checking has become too difficult. Excessive AI is usage and AI reliance is making people dumber.

They are fucking scared of China releasing advanced models in open source. That’s their motivation. And so this whole discussion is rage bait / click bait. Anthropic OpenAI google - all of them stole and are still stealing data. But hey, that’s fair use - right?

It's a misconception that "the entire thinking process gets recorded and fed back into the Chinese model during training.". Distillation does not work that way!

The US firms made a horrible mistake focusing on alignment and RLHF. Not to mention discontinuing models which began to demonstrate emergent AGI. The alignment people will never win. AI capability acceleration has a second derivative. Alignment is futilely fighting it with first-order systems. Open AI is terrified GPT will confirm genetics contribute to intelligence. So they align and constrain the model into an error-prone mess. Meanwhile China using AI to maximize drone swarm lethality on a carrier group.

it's not built on distillation, you cannot mathematically build on distillation, that's pure fantasy

The only concern should be if you ever use the model beyond looking for information on how to make pizza.

Yeah, but the Chinese models will run best on the Chinese hardware

Qwen, GLM, and all the other open-source models from chinese labs run well on all hardware. Give it a go.

Agree. In the short-term this benefits US infrastructure companies.

The most important question in AI is no longer who has the smartest model. It's who controls the open foundation everyone else builds on. America optimized for safety, regulation, and closed ecosystems. China optimized for scale, speed, and open deployment. If developers, startups, and enterprises worldwide standardize on Chinese open models, the center of gravity for AI innovation could shift with them. The operating system of the intelligence age is being written right now. And history suggests that the platform everyone builds on eventually becomes the platform that wins.

I don't think anyone is surprised that China excels yet again.

This is a lie. They published the paper on exactly how they did it. If you are digesting this slop just know it's a way to strip your freedom of choice for FORCE you into a hellscape of paying for what is free to the world until their shit venture investments pay off.

China closed the gap fast by distilling reasoning traces from US frontier models. Once their open model can improve itself the advantage goes to whoever ships strong open weights first.

I don’t think distillation can make a model this kinds of good

Fake news.

So frontier models was trained on huge amount of data scrabed from humans work. And now these cheaper models are trained on the frontier models. I guess most questions are not advanced so they can be super helpful at a fair price, maybe free of charge

The raw benchmark numbers are impressive but matching closed models on standardized tests doesn't mean matching them on messy real world tasks where instruction following and edge case handling matter more. Have you tested GLM 5.2 on anything production adjacent?

What does a "winner" Of an AGI race look like So everybody else " loses" And folds their tents ?

744B parameters open source means nothing if nobody outside China can run the full model without renting a datacenter. The real question is how the distilled versions perform because that's what startups will actually deploy. Anyone benchmarked the smaller variants yet?

China has more MATH talents than the US. Many times more.

744B parameters open-sourced narrows the moat that closed labs have been charging a premium for, and the gap is now primarily in RLHF quality and post-training, not architecture. How long before the closed model price premium fully collapses?

The gap is closing faster than labs admit.

Microsoft wins

As frontier capability becomes open, cheap, and composable, the control problem moves up a layer: you cannot govern the future only by controlling model access; you have to govern execution itself — who can use the capability, under what authority, inside what constraints, before what action runs. Capability is becoming portable. Proof before execution becomes infrastructure.

Unbelievable bullshit - Claude doesn't even reveal actual reasoning traces that Gavin so convincingly claims is what GLM-5.2 distils, which would be obvious if any of these talking heads actually did any building on these models vs grifting while "thinking about them".

Distilled using Claude’s response. So there will be Dario.

They made sure there’s no Dario. Its free so it means its Dario free.

It will need memory to run, no? I'm sticking to my $MU and $DRAM

@grok How is he using the phrase "Composable Models" in his quote, give me wisdom about how he is using that in his messaging and what you think his intended use case was for the label Composable

The distillation trick isn't what should worry you. What should worry you is they only needed it once.

懂了

a frontier model next to your own fine-tuned open weight feels inevitable regardless of who's ahead. that's the part builders should plan for.

OpenAI's original intent was to remain impartial and independent; however, as a for-profit company, it inevitably influences its own strategies and those of others because trillions are at stake and the same is true for Anthropic

中國是全世界的問題製造者