The current balance of power in open models
4 hours ago
- Open-weight models have public weights and inference code, while open-source models also include training code and data; closed models are API-only.
- Chinese open-weight models (e.g., Qwen, GLM, Kimi) have surpassed American counterparts in downloads and benchmark performance since mid-2025, leading by ~1.6B Hugging Face downloads and higher AAII scores.
- American open-weight models lag behind Chinese models by 6–9 months relative to the closed frontier (OpenAI/Anthropic), while Chinese models are only 2–5 months behind.
- Distillation from closed American APIs accounts for only about 1–2 months of the Chinese lead, not the entire gap.
- Chinese open-weight models dominate adoption in academia (38% of arXiv papers mentioning a Chinese model vs. 28% for U.S.) and are widely used by major companies like Harvey, Cursor, DoorDash, Airbnb, and Perplexity.
- Open model usage is exploding in high-value industries, with platforms like OpenRouter processing ~80T tokens per week, of which >80% are from Chinese models.
- Risks from advanced open-weight models (e.g., cybersecurity) require ecosystem-level preparation, but restricting access would hurt American businesses more than adversaries.
- The U.S. must reinvest in open models to regain leadership in AI research and economic diffusion, as open models are becoming the substrate for broad access to transformative intelligence.