I ran a similar analysis on 221 MATS alignment and safety projects.
In 2026, 72% of projects using named models use a Chinese open-weight family, versus 54% using an American one. Qwen appears in 66%, ahead of Llama at 44% and DeepSeek at 32%.
The interesting difference is that closed models still appear in 90% of these projects, often as judges, monitors, or frontier baselines.
Chinese open models are becoming the experimental substrate of alignment research, while closed American models remain key evaluation tools.
Over the weekend I had Codex parse 500K arXiv AI/ML papers since ChatGPT to understand which open models are used for research. In 2024, ~30% of papers mentione...