Chinese open-weight AI models now account for more than 80 percent of weekly token volume on OpenRouter, up from roughly 70 percent a year ago, according to AI researcher Nathan Lambert. He adapted the analysis from remarks he delivered to Congressional members and staff, a public version of testimony originally prepared for lawmakers.
Lambert previously helped build Olmo, an open-source model line from the Allen Institute for AI, and now tracks the field through a project he calls American Truly Open Models, or ATOM. He lays out the shift across three separate measures rather than one headline stat. On Hugging Face downloads, China overtook the US in July 2025, driven largely by Alibaba’s Qwen family. Its cumulative lead has since grown to about 1.6 billion downloads, with total Chinese downloads reaching 3.2 billion, twice the American total.
On capability benchmarks, the pattern repeats. Lambert’s own Artificial Analysis Intelligence Index puts three Chinese systems on top: Moonshot AI’s Kimi K3 scores 44, Z.ai’s flagship GLM-5.3 scores 45, and the smaller GLM-5.3-Flash scores 42. Nvidia’s Nemotron 3 Ultra trails at 23, while Thinking Machines’ two entries, Inkling and Inkling Small, tie for the best American result at 26. Lambert notes these are his own dashboard’s figures, not an independent third party’s, but the direction is consistent with the download data.
Real-world adoption tells the same story. Building AI features on Chinese open models such as Kimi, Qwen, and DeepSeek is now common practice: Perplexity relies on DeepSeek, Airbnb uses Qwen, and DoorDash has adopted Kimi, according to Lambert. He also scanned five major arXiv research categories and found any Chinese open model now appears in over 40 percent of papers, versus roughly 30 percent that cite an American open model, with the Chinese share growing faster.
Lambert is careful to separate two different competitions. Chinese open-weight models trail the closed American frontier (OpenAI, Anthropic) by an estimated 2 to 5 months, and that gap has been narrowing for three years, he writes. American open-weight models trail that same closed frontier by 6 to 9 months and also lag the Chinese open models, meaning the real story is not China catching the US, but American open models falling behind on two fronts at once. Lambert adds one caveat: the Chinese lead is narrowest on tasks with clear commercial demand, like agentic coding, and widest on open-ended scientific work such as physics or biology.
He also addresses the most common explanation for Chinese progress: distillation, the practice of training a model on a stronger model’s outputs. Lambert estimates that if distillation were fully blocked, for instance through know-your-customer verification at Anthropic and OpenAI, the capability gap would only widen by 1 to 2 months. That is a narrow band for a technique often cited as the whole explanation.
The policy tension Lambert raises is a genuine bind rather than a rhetorical one. Restricting the strongest Chinese open models over security concerns would primarily hurt the American startups and enterprises that already depend on them, since open weights cannot be recalled once released. He points to Hugging Face’s own use of a Chinese model to analyze a cyberattack, after closed American models declined to help, as a case where restriction would have blocked a defensive use.
This is Lambert’s own framing of a fast-moving, contested landscape, not a neutral survey, and his dashboard methodology, built on Hugging Face downloads, OpenRouter share, and arXiv mentions, is one lens among several plausible ways to measure “leadership.” Enterprises weighing which open-weight stack to standardize on should treat vendor lock-in around Chinese-model APIs as a live procurement risk, not just a cost line, given how much of the current API surface for those models still runs through third-party inference platforms rather than the labs themselves.
Nathan Lambert published this analysis on Interconnects on September 21, 2026.
