The current balance of power in open models
Chinese open‑weight models now outpace U.S. ones in downloads, benchmarks, and industry use, reshaping the AI landscape.
Why Now
The article reports that since mid‑2025 Chinese labs have overtaken U.S. labs in open‑weight model performance and adoption, highlighted by download stats and benchmark scores.
What Happened
Chinese open‑weight models such as GLM‑5.3, GLM‑5.3‑Flash, and Kimi K3 lead the Artificial Analysis Intelligence Index with scores 45–44, while top U.S. models score 23–26. Downloads on Hugging Face show China’s lead grew to 1.6 B out of 3.2 B total, double the U.S. share. Open‑weight models are widely used by U.S. companies (e.g., DoorDash, Airbnb) and in academia, with Chinese models cited in 30% of recent arXiv papers.
Why It Matters
This shift gives Chinese labs a competitive edge in agentic coding and other high‑impact tasks, while U.S. firms increasingly rely on Chinese models for cost‑effective AI features. It also raises regulatory and security concerns, as open‑weight models are harder to control and can be used by bad actors.
The Limitation
The analysis relies on public download and benchmark data, which may not fully capture private deployments or the true quality of models across all tasks.
What You Can Do
Benchmark your own workloads against Chinese open‑weight models to assess performance gaps and consider sourcing models from U.S. open‑source projects for greater control.
Source
Read original sourceWhy we picked this
Discussion of open models and congressional testimony, relevant policy news.