Congressional Briefing: China Has Led US in Open-Weight AI Since 2025, Gap Widening
Summary
- • Chinese open-weight AI models have led US counterparts since April 2025 in capabilities and downloads
- • Top Chinese models score 42-45 on AAII benchmarks; top US open-weight models score only 23-26
- • Chinese models cited in 40%+ of AI research papers; Alibaba Qwen alone cited in 30% vs. Llama at 21%
- • Harvey, Cursor, DoorDash, Airbnb, and Perplexity all building commercially on Chinese AI models
Details
China leads Hugging Face downloads by 1.6B; China total 3.2B vs. US ~1.6B
Since China took the Hugging Face download lead in July 2025, primarily through Alibaba's Qwen models, the gap has widened to 1.6 billion more downloads. China's total is approximately 3.2B vs. the US's ~1.6B, tracked by the author's ATOM Project tools.
Top Chinese AAII scores: GLM-5.3 (45), Kimi K3 (44); US open-weight best: 26
On the Artificial Analysis Intelligence Index as of September 2026: top Chinese open-weight models are Z.ai GLM-5.3 (45), Moonshot Kimi K3 (44), and GLM-5.3-Flash (42). Top US open-weight models are Thinking Machines Inkling (26) and Nvidia Nemotron 3 Ultra (23) — released June-July 2026.
Chinese models in 40%+ of AI papers; Qwen at 30% vs. Llama at 21%
A scan of 5 major arXiv ML categories (cs.AI, cs.CL, cs.CV, cs.LG, stat.ML) shows Chinese open-weight models cited in 40%+ of recent AI papers. Alibaba Qwen alone is at 30%, surpassing Llama's 21% — remarkable given that Llama peaked at ~23% in fall 2024.
Harvey, Cursor, DoorDash (Kimi), Airbnb (Qwen), Perplexity (DeepSeek) use Chinese models
Major US companies have adopted Chinese open-weight models commercially. Chinese models represent 80%+ of OpenRouter open-model usage (up from ~70% in 2025) and ~95%+ on open-source coding agent OpenCode. A growing wave of startups is entering enterprise agreements with Chinese AI labs.
Chinese models 2-5 months behind closed US frontier; US open-weight 6-9 months behind
Chinese open-weight models trail OpenAI/Anthropic's closed frontier by approximately 2-5 months. American open-weight models trail that frontier by 6-9 months — meaning US open models are significantly behind both Chinese open-weight and US closed models on a capability basis.
Distillation explains only 1-2 months of Chinese open-weight advantage; rest is genuine
Estimates suggest that fully preventing API distillation (e.g., via KYC tools at Anthropic and OpenAI) would narrow the Chinese open-weight performance gap by only 1-2 months. The majority of China's advantage reflects genuine technical progress, faster release cadence, and a narrower task distribution at release time.
Policy recommendation: invest in US open models for soft power and risk coordination
The briefing argues that US ownership of competitive open models enables better coordination of global AI risks and accelerates domestic AI diffusion. Framed as both a security and soft power imperative — the alternative is American businesses increasingly dependent on Chinese AI infrastructure.
Briefing delivered to Congressional members and staff by Interconnects AI researcher
The analysis was prepared for a Congressional briefing on open-weight AI in the US-China competition lens and published publicly. The author previously spent 2.5 years at the Allen Institute for AI working on open models and created the American Truly Open Models (ATOM) Project in August 2025.
US-China open-weight AI competition — capability benchmarks, download metrics, academic adoption, and policy recommendations
What This Means
This congressional briefing paints a concerning picture for US open-weight AI competitiveness: Chinese labs have not only led since 2025 but are entrenching that lead through dominant adoption in academia, commercial products, and global developer ecosystems. With top US open-weight models scoring nearly half what Chinese counterparts achieve on capability benchmarks, and major US companies like Harvey, Cursor, and DoorDash already depending on Chinese AI infrastructure, the challenge goes far beyond a simple leaderboard gap. The distillation-from-American-models explanation covers only 1-2 months of the differential, meaning the gap is largely structural. The policy prescription — invest in US open models as national infrastructure — reflects growing concern that America's AI advantage may be limited to a handful of closed frontier labs while the open ecosystem, which underpins global AI research and commercial adoption, becomes Chinese-dominated.
Sources
- The current balance of power in open modelsInterconnects