Chinese Open-Weight AI Models Now Dominate Research and Industry

Chinese models capture majority market share

Chinese AI companies now control the open-weight language model ecosystem, according to testimony prepared for Congress by AI researcher Nathan Lambert. Chinese models account for 3.2 billion downloads on Hugging Face—double the U.S. total—and power over 80% of inference volume on platforms like OpenRouter, up from 70% in September 2025.

The shift represents a reversal from early 2024, when Meta's Llama models dominated both research and commercial applications. Today, Alibaba's Qwen family and models from Moonshot AI and Z.ai have become the foundation for AI development across academia and industry.

On the Artificial Analysis Intelligence Index, the top three open-weight models are all Chinese: GLM-5.3, GLM-5.3-Flash, and Kimi K3, scoring 45, 42, and 44 respectively. The leading U.S. models—Thinking Machines' Inkling and Nvidia's Nemotron 3 Ultra—score between 23 and 26, representing a performance gap of approximately 2-5 months behind Chinese open-weight offerings and 6-9 months behind closed American frontier models like GPT-4 and Claude.

Academic research tilts toward Chinese foundations

The adoption gap extends deep into research infrastructure. Lambert's analysis of arXiv preprints shows Chinese open-weight models mentioned in 38% of AI papers, compared to 28% for U.S. models. Qwen alone appears in 30% of recent papers, surpassing Llama's 21% despite Meta's model maintaining remarkable longevity since its 2023 release.

Major U.S. technology companies have begun building products on Chinese model foundations. Harvey's legal agent, Cursor's coding tools, and enterprise features at DoorDash, Airbnb, and Perplexity all rely on Chinese open-weight models. Many startups are now entering licensing agreements with Chinese labs—a form of cross-border AI collaboration Lambert notes he hasn't witnessed previously in his career.

The Chinese labs achieve competitive performance despite fewer resources through faster release cycles and narrower task focus. They've also shifted strategy in 2026, moving from in-house data workflows to purchasing cutting-edge training data from both American companies and Chinese startups.

Distillation explains only part of the gap

While Chinese labs do use distillation—training models on outputs from stronger American systems—to accelerate development, Lambert estimates this technique accounts for only 1-2 months of their competitive advantage. Even with full prevention of distillation through know-your-customer tools at OpenAI and Anthropic, Chinese models would remain close to the frontier.

U.S. models show promise when they achieve comparable capabilities. OpenAI's first open-weight release since ChatGPT, gpt-oss, became one of the most adopted models ever. Google's Gemma 4 matches adoption rates of Qwen's popular small models, and Nvidia's Nemotron maintains modest usage despite more capable alternatives.

Why it matters

The dominance of Chinese open-weight models creates strategic dependencies for American businesses and researchers. When HuggingFace needed to analyze a recent cyberattack, they used Chinese models because closed American systems wouldn't answer their security queries. As open models approach frontier capabilities and enter high-value industries like software engineering and legal services, the infrastructure layer of AI development is shifting away from U.S. control. This affects not just commercial competitiveness but America's position as the center of AI research—a shift Lambert describes as accelerating due to both technical factors and the closed nature of leading American labs.

These details were first reported by Nathan Lambert in testimony prepared for Congressional briefings on U.S.-China AI competition, published on Interconnects AI.