Chinese AI models: close to the frontier, far from the compute

THE THESIS. China has not reached the AI frontier. It has achieved something that may matter more: it has made near-frontier AI cheap, open and Chinese. There is still a gap with the best models from Anthropic and OpenAI, but it's measured in months, not years. Talent isn't what holds China back. Compute is: the number of chips available to train models and serve them to users. Our take is cautious optimism. In the short term the risks are real: technological dependence, censorship inside the models, and a trade war that splits the market in two. In the long term, powerful models that cost next to nothing are part of the abundance we argue for.

THE REAL GAP. On October 27, Artificial Analysis released version 5 of its Intelligence Index. It's a score that combines many tests, and the industry follows it closely. What's new is that 60% of the tests now use private questions. Labs can't have seen them during training. The result was a cold shower. The firm itself says the gap between the best closed model and the best open-weight model grew from 3 points to 12. An open-weight model is one that anyone can download and run on their own servers.

Moonshot's Kimi K3.1 Pro fell from fourth place to ninth. Coverage of the update says it dropped from 55 to 46 points. Even so, it's still the best Chinese model. The best closed models from Anthropic and OpenAI score around 57-58. Open models that are already available for download sit at about 32. That includes Zhipu's GLM-5.3 and DeepSeek V4.1 Flash. On the hardest tasks the gap grows. According to an analysis cited by the tech press, the best closed models solve about 60% of the tasks in Terminal-Bench-Science. Most Chinese open models solve around 10%.

Other measurements show a smaller gap. Stanford University's 2026 AI Index puts it at 2.7% between the best US model and the best Chinese one. The Institute for Progress (IFP), a Washington think tank, estimates that the US is about 6.5 months ahead.

Our take: both are true at once. On public, well-known tests, China is nearly level. On new, private tests built around long tasks, a real Western lead shows up. That backs a view we have held for months: claims that a model 'matches X' are usually marketing until they're checked on hard tests. Part of the closeness on public tests probably comes from tuning models to those tests.

WHO'S WHO. DeepSeek released V4 and V4-Pro on April 24, 2026, under the MIT license, one of the most permissive there is. They are mixture-of-experts (MoE) models: only part of the network switches on for each answer, which makes them cheaper to run. The company reported 80.6% on SWE-bench Verified, a test of whether a model can fix real bugs in code. That number comes from the company itself, not an independent evaluator. Zhipu, now called Z.ai, released GLM-5 in February and GLM-5.2 in June. GLM-5.2 was the first open model to top the previous version of the Artificial Analysis index, with 51 points. Moonshot launched Kimi K3 in July. It has 2.8 trillion parameters, and VentureBeat calls it the largest open model released so far.

Alibaba plays a different game. Its Qwen family has passed 1 billion cumulative downloads on Hugging Face, the main platform for open models. Thousands of derived models are built on it. Its flagship, Qwen3.8-Max, is only offered through an API, meaning as a pay-per-use service. MiniMax and StepFun round out the leading pack.

Our take: what stands out isn't any one champion. It's how many labs are competing. China has five or six that can release a near-frontier model every few months. That competition at home pushes prices down for everyone.

WHY CHINA GIVES ITS MODELS AWAY. According to Hugging Face's Spring 2026 report, Chinese models made up 41% of downloads between February 2025 and February 2026. US models made up 36.5%. Figures for actual use vary a lot by source. A study by a16z with OpenRouter, a platform that offers many models through one API, found an average 13% share of traffic over a year. A Mozilla analysis puts it at 45% in April 2026. The two cover different periods, so they can't be compared directly.

Price explains much of this. In May, DeepSeek cut the price of V4-Pro by 75%. It now costs $0.435 per million input tokens and $0.87 per million output tokens. A token is roughly three-quarters of a word. At launch it cost $1.74 and $3.48. Kimi K3 is more expensive: $3 for input and $15 for output per million tokens. Even so, Artificial Analysis calculated that each task on its previous index cost $0.94 with it.

Our take: releasing the weights is a strategy, not charity. With less compute, China can't win by selling the smartest model. It can become the standard others build on: startups, governments and universities that don't want to depend on a US API. It's a way of exporting tech independence. But there's fine print. Chinese models are subject to Beijing's content rules. And Anthropic says DeepSeek used queries to Claude to learn how to steer conversations away from topics censored in China.

There is also an open accusation. In February, Anthropic accused DeepSeek, Moonshot and MiniMax of industrial-scale distillation. Distillation means training your own model on the answers of someone else's model. According to Anthropic, there were more than 16 million exchanges with Claude through about 24,000 accounts. OpenAI made a similar accusation against DeepSeek to the US Congress. The three Chinese companies did not respond at the time. Several critics pointed out that US labs also face disputes over their own training data. If the accusations are confirmed, they would explain part of China's speed, but not all of it.

THE BOTTLENECK: CHIPS AND MEMORY. The timeline shows that both countries are restricting trade. In April 2025, the US began requiring a license to sell China the H20, a cut-down Nvidia chip. In September 2025, according to the Financial Times, China's internet regulator told ByteDance and Alibaba to stop testing and ordering the RTX Pro 6000D, another Nvidia chip designed for China. In January 2026, Washington allowed sales of the H200 in exchange for keeping 25% of the revenue. But in the quarter that ended July 26, those sales were less than 1% of Nvidia's data center revenue, according to Bloomberg. The company blamed this partly on objections from Beijing.

China's alternative is Huawei. According to the company and press reports, GLM-5 was trained on about 100,000 Huawei Ascend 910B chips, with no Nvidia hardware. No one has verified this independently. Huawei has confirmed that its new Ascend 950 can run DeepSeek V4. But the scale is very different. Epoch AI, an institute that tracks AI progress, estimates that Huawei will produce less than 4% as much AI compute as Nvidia in 2026. The main limit isn't SMIC, China's largest chipmaker. It's HBM, the high-speed memory that sits next to the chip and feeds it data. In June, the research firm SemiAnalysis estimated that only about 25% of the eight-layer HBM3 made by CXMT, China's leading memory maker, came out usable.

The US compute lead is large, but its size depends on what you measure. IFP estimates the US has about 16 times more capacity for training models. According to Epoch AI data, the US holds nearly three-quarters of the world's AI compute and China just over 14%. SemiAnalysis, cited by the Financial Times, puts operational AI data center capacity at 56 gigawatts in the US and 24 in China. And according to Stanford, private AI investment was 23 times higher in the US: $285.9 billion versus $12.4 billion.

Our take: what's surprising is how little this shows up in the models. With about 16 times less training compute and 23 times less private investment, China is months behind. That tells us two things. Efficient engineering pays off a lot. And following a leader is cheaper than breaking new ground. But compute matters most in the newest areas: agents that work for hours and long reasoning. That's exactly where index v5 shows the gap. It all fits.

CONTROLS, ENERGY AND CHINA'S RESPONSE. China has one clear advantage: energy. According to Al Jazeera, it generates more than twice as much electricity as the US. According to Washington Post calculations cited by China Daily, its data centers pay about 3 US cents per kilowatt-hour, a third of US prices. Bernstein expects more than 5 trillion yuan of investment in China's power grid between 2026 and 2030. Chips are scarce in China. Electricity is not.

Technology is now a bargaining chip. According to press reports, Chinese negotiators are linking sales of rare earth magnets, which are vital to the defense industry, to US restrictions on chips. The pause on China's toughest rare earth controls expires on November 10, and reports conflict on whether it has been extended. On October 16, Beijing also published draft export controls on its own AI: Announcement No. 48 from the Ministry of Commerce (MOFCOM). It covers the weights of models that exceed 10^26 training operations or 2 trillion parameters, or that rank in the top ten on certain tests. Anything already published is excluded. Offering a model through an API does not count as an export. Analysts disagree on whether openly releasing a future model above those thresholds will require permission. Public comment closes on November 15. Kimi K3, at 2.8 trillion parameters, already exceeds one of those thresholds.

Our take: US controls have slowed the amount of compute China has, not what it can do with it. They have delayed volume, but they have sped up China's own industry (Huawei, CXMT) and pushed Beijing to turn away US chips. That's what we have been arguing: controls push Chinese self-reliance forward more than they hold it back. What's new is that Beijing is starting to treat its model weights as a strategic asset, just as Washington did with Mythos. The era of Chinese open weights could become conditional.

IMPLICATIONS. For companies and developers, Chinese open models are a real alternative on cost and control. They can run on your own servers without sending data to China. But they need auditing for content bias, security and regulatory compliance. In geopolitics, the leaderboard matters less than who sets the standards others build on. We said it in the Google-Microsoft case: whoever controls the plumbing wins, not just whoever has the smartest model.

In the short term, the risk is a split into two incompatible ecosystems, with separate chips, models and rules. In the long term, competition between labs keeps driving the cost of intelligence down. When a near-frontier model costs less than a dollar per million tokens, a hospital or a university without a big budget can use it to research diseases. That's the path to the abundance we argue for, as long as governance keeps up.

What to watch in the coming weeks: whether Moonshot finally releases the Kimi K3.1 Pro weights (it announced them for November 4, and we have not been able to confirm the release); the final text of Announcement No. 48; how the rare earths standoff plays out after November 10; and the production pace of the Ascend 950 and of Chinese HBM memory. Those will decide whether the gap closes or widens.