By mid-2026, Chinese AI labs have taken a clear lead in the open-weight language model ecosystem, surpassing U.S. counterparts on downloads, benchmark scores, and academic usage. Open-weight models are becoming economically significant across industries and research, while also creating new safety and policy challenges that technical restrictions alone struggle to address.
Open-weight vs. open-source vs. closed models
- Open-weight language model: weights are publicly available for inspection or downstream use; examples include Meta’s Llama, Alibaba’s Qwen, and Google’s Gemma. These releases typically include licenses and inference code (e.g., Transformers, VLLM, SGLANG).
- True open-source model: in addition to weights, the full reproduction materials (training code and training data) are available. Prominent open-source efforts have been U.S.-led, including Allen Institute for AI’s Olmo models, OpenAthena’s Marin, and EleutherAI’s Pythia.
- Models lie on a spectrum: some releases (e.g., Nvidia Nemotron) share much training data and are relatively open without being fully reproducible. Closed models (API-first offerings like GPT-4 or Claude) likewise vary by how much internal detail they reveal.
State of competition: how China pulled ahead
- Downloads and adoption: Alibaba’s Qwen family helped China take the Hugging Face lead in July 2025. According to the American Truly Open Models (ATOM) Project tracker, as of September 14, 2026 Chinese models lead by about 1.6 billion downloads, totaling ~3.2 billion — roughly twice U.S. totals.
- Benchmark performance: on the Artificial Analysis Intelligence Index (AAII) the leading Chinese open-weight models — Z.ai’s GLM-5.3 and GLM-5.3-Flash, and Moonshot AI’s Kimi K3 — scored 45, 42, and 44 respectively. Leading American open models (Thinking Machines’ Inkling and Inkling Small) scored 26, and Nvidia’s Nemotron 3 Ultra scored 23. The top U.S. open models were released in June–July 2026 and are updated less often than many Chinese releases.
- Capability timing: Chinese open-weight models are roughly 2–5 months behind the closed American frontier models, while American open-weight models are about 6–9 months behind the likes of OpenAI and Anthropic.
Factors behind China’s strength
- Faster release cadence and narrower task focus: quicker releases and optimization for a slightly narrower distribution of tasks can boost public benchmark performance at release time.
- Data strategies: in 2026 leading Chinese labs (e.g., Moonshot AI, Z.ai) shifted from building all data workflows in-house to purchasing cutting-edge training data, including challenging RL environments for agentic tasks, from both U.S. firms and Chinese startups.
- Distillation: Chinese companies widely use distillation (training models on outputs of stronger models). The author estimates stopping distillation (e.g., via KYC at Anthropic/OpenAI) would only widen the gap by ~1–2 months, so distillation is not the sole explanation for Chinese success.
Adoption across platforms, industry, and academia
- Platform usage: OpenRouter — an inference platform that lets users switch among open and closed, U.S. and Chinese models — grew from ~1T tokens/week for open models in September 2025 to ~80T tokens/week by 2026. Chinese models’ share rose from ~70% to over 80% in that period. OpenCode and similar commercializers report inference volumes where Chinese models account for ~95% or more.
- Corporate and startup use: multiple notable companies build on Chinese open-weight models (examples: Harvey, Cursor, DoorDash’s use of Kimi, Airbnb’s use of Qwen, Perplexity’s use of DeepSeek). Increasingly, U.S. startups and companies enter enterprise agreements with Chinese labs to license models for product use — a new form of cross-border collaboration.
- Academic adoption: scanning arXiv papers in five popular ML categories (cs.AI, cs.CL, cs.CV, cs.LG, stat.ML) shows mentions of any open model rose from ~2% in January 2023 to ~50% in September 2026. Llama mentions peaked around 23% in fall 2024 and were ~21% by September 2026; Qwen mentions rose to ~30%. Overall, mentions of Chinese open-weight models exceed U.S. models (~40% vs. ~30%), and China’s share is growing.
Risks and regulatory challenges
- Difficulty of restricting access: open-weight releases are hard to keep out of malicious hands once distributed. An example: HuggingFace used a Chinese open-weight model to investigate a cyberattack because closed models would not answer their queries.
- Regulatory uncertainty: incidents and documentation about frontier-model risks (e.g., the OpenAI–HuggingFace incident) have increased scrutiny, but how to legally and effectively regulate open-weight releases remains unclear.
- Ecosystem preparation as policy: because outright restriction is impractical, the author argues for investing in domestic open models and preparing the broader ecosystem for mitigation and coordinated response.
Summary conclusions
- In 2026 Chinese labs are the clear leaders of the open-weight AI ecosystem; open-weight models have crossed an inflection point in economic relevance while U.S. labs are increasing activity but have not closed the gap.
- Distillation contributes to Chinese progress but does not fully explain it; faster releases, focused tasks, and evolving data strategies also matter.
- Open-weight models are enabling rapid diffusion of AI into high-value sectors (software engineering, legal, financial services) and creating new capability-driven risks, particularly in cybersecurity, that demand ecosystem-level responses.
- The recommended approach is to support U.S. investment and ownership of open models to improve coordination on risk management, while engaging internationally to understand how open models are used beyond the U.S. and China.
Acknowledgements: Florian Brand and Kevin Xu provided feedback and suggestions. For more data, see the Interconnects Dashboard and the open-source AI reading list.
