Chinese Open-Weight Models Now Lead US Rivals by Wide Margin, Interconnects Analysis Finds

The Gap Has Widened, Not Closed

Chinese open-weight AI models now clearly outperform their American counterparts on both adoption and capability benchmarks, according to a detailed analysis by AI researcher Nathan Lambert, published as an expanded version of testimony he prepared for a Congressional briefing on U.S.-China AI competition.

As of September 14, 2026, the top three open-weight models by Artificial Analysis Intelligence Index (AAII) score are all Chinese: Z.ai's GLM-5.3 (45), Moonshot AI's Kimi K3 (44), and GLM-5.3-Flash (42). The leading American open-weight models — Thinking Machines' Inkling and Inkling Small, and Nvidia's Nemotron 3 Ultra — score 26, 26, and 23 respectively. Lambert notes the top American open models were released in June and July of 2026 and are updated less frequently than Chinese competitors, whose releases have consistently beaten American scores by 2-6 months, citing GLM-5 and DeepSeek V4 Pro as earlier examples.

Adoption metrics tell the same story. China overtook the U.S. in Hugging Face downloads in July 2025, driven largely by Alibaba's Qwen models. Since Lambert launched his American Truly Open Models (ATOM) tracking project in August 2025, China's download lead has grown to roughly 1.6 billion, with a cumulative total of 3.2 billion downloads — double the American total.

Where the Gaps Sit

Lambert estimates Chinese open-weight models trail the closed American frontier (OpenAI, Anthropic) by roughly 2-5 months, while American open-weight models trail the same frontier by 6-9 months. Chinese labs are closest to parity in agentic coding, where commercial demand is highest, and furthest behind on open-ended scientific reasoning tasks like physics and biology.

Distillation from closed American APIs is often cited as an explanation for Chinese progress, but Lambert argues its effect is limited: he estimates that even if labs like Anthropic and OpenAI fully blocked distillation with know-your-customer verification, the capability gap would widen by only 1-2 months. He also notes Chinese labs including Moonshot AI and Z.ai shifted through 2026 from building training data in-house to purchasing challenging agentic RL environments from both American and Chinese data vendors.

Policy Tension

The analysis flags a structural problem for restricting open-weight releases on security grounds: once weights are public, there is no reliable mechanism to keep them from any actor, and restricting access would primarily harm American businesses already dependent on them. Lambert cites a recent incident in which Hugging Face used a Chinese open-weight model to analyze a cyberattack because closed American models declined to answer the relevant queries — illustrating, he argues, that U.S. companies already lean on Chinese open models when American alternatives refuse.

What This Means

The data points to a durable shift rather than a temporary lead: Chinese labs are releasing more frequently, scoring higher on independent benchmarks, and being downloaded at twice the rate of American open models. New entrants like Arcee AI, Poolside, and IBM are expanding the American open-weight field but not closing the performance gap. For U.S. policymakers, the report's core argument is that restricting Chinese open-weight access would likely harm American developers who already depend on these models more than it would slow Chinese AI progress — making continued investment in domestic open-weight development, rather than restriction, the more viable lever.