Lambert Reports a Chinese Lead in Open-Weight Model Downloads

Nathan Lambert prepared a briefing for members of Congress on where open-weight AI models stand today. Open-weight models are AI systems whose underlying files get published for anyone to download and run on their own hardware. That’s different from a closed model, which you can only reach through a company’s paid app or API, a set of instructions a program uses to talk to a service over the internet. His conclusion: Chinese labs have not just caught up. They have pulled ahead, and American companies are already building on their models.

What he said

Lambert writes that the shift already happened, not that it’s coming. “Chinese AI companies are the clear leaders in open weight models,” he states. He puts a date on when that lead opened up: “Chinese open-weight models surpassed American open-weight models in these two key areas about 18 months ago.”

He backs that with download and usage data. “China’s download lead has grown to about 1.6B, with a total of 3.2B downloads, twice that of America’s total,” he writes. On where developers are actually sending API traffic today, he points to OpenRouter, a marketplace that routes requests across many AI providers. His comparison concerns tracked open-model traffic on that platform, rather than all AI usage or all OpenRouter traffic. “Chinese models have grown from ~70% market share to over 80% of usage,” he writes, discussing the open-model usage series.

The adoption claim is the one that should get builders’ attention. “Many prominent technology companies and startups have been building on Chinese open-weight models for their AI features, such as Harvey, the legal agent, Cursor, the coding agent, and DoorDash’s use of Kimi models, Airbnb’s use of Qwen, or Perplexity’s use of DeepSeek,” Lambert writes.

Who he is and why the take carries weight

Lambert writes Interconnects, an independent newsletter that tracks AI model releases, training methods, and industry strategy. Before that, he spent two and a half years at the Allen Institute for AI (Ai2). He led post-training research there, the work of fine-tuning an already-built model’s behavior after its initial training, and built the Olmo and Tulu open model families. He left this past June to work independently. He’s also the author of a book on reinforcement learning from human feedback, the training technique behind how today’s chatbots learn to follow instructions.

Lambert adapted this piece from remarks he prepared for a congressional briefing. He distinguishes open-weight models, whose trained files a company publishes under a license while keeping training data and code private, from fully open-source models, which publish the training data and code too. Most of what gets called “open” today, including the Chinese models he’s describing, is open-weight, not open-source.

What he gets right, and where it’s incomplete

The download and usage numbers are Lambert’s own count from public platforms. They support his account of adoption on those platforms, but do not capture private deployments or providers that withhold model-level usage. The adoption list, naming Cursor, Perplexity, DoorDash, and Airbnb by name, is the strongest part of the piece. It shows the shift isn’t confined to hobbyists comparing benchmark scores. It’s already inside products people use.

Where the piece is thinner is on causes and endpoint. Lambert discusses release speed, task focus, work cultures, training data, and distillation, while acknowledging debate over the causes. His benchmark comparison also has a limit worth naming. He cites Chinese models Z.ai’s GLM-5.3 and Moonshot AI’s Kimi K3 scoring higher than the American open models he names. That comparison uses the Artificial Analysis Intelligence Index, a third-party benchmark that combines several tests into one capability score. The comparison he presents here is among selected open models; it does not establish that they outperform closed frontier models. Lambert separately estimates a gap of two to five months between Chinese open models and the American closed frontier, with the gap varying by task. His closing recommendation, that the US should “continue to enable investment in open models,” is his policy prescription for Congress.

Why it’s notable

Most coverage of the US-China AI race focuses on frontier, closed models: which lab has the smartest chatbot. Lambert’s briefing is about a different, more practical layer. It’s about which models are cheap and available enough for a startup to build on, run on its own servers instead of paying a vendor for cloud access, or fine-tune for a narrow task. That’s the layer where cost and control matter most for builders. By his numbers, China has taken a clear lead there.

What it means for builders

Cost, license terms, hardware requirements, and task performance all affect the model choice. Lambert identifies Chinese labs such as Z.ai, Moonshot AI, and Alibaba’s Qwen team as important suppliers, and cites US companies using their models. Download counts and benchmark scores alone do not establish the cheapest model to self-host.

That doesn’t mean every builder should switch. It means the choice deserves the same evaluation you’d give an American vendor. Check the license terms for commercial use. Test the model against your actual task instead of a general benchmark. Include data-handling requirements, deployment location, and applicable organizational restrictions in the comparison. Lambert’s account makes open-model selection a practical product decision alongside the policy debate.

Updated October 13, 2026: the reported OpenRouter share concerns tracked open-model usage. Lambert’s benchmark comparison covers selected open models, while his separate estimate of the gap to closed models varies by task.

End of article