Xiaomi releases MiMo-V2.6-Pro-RL, a 1.02T/42B open omnimodal model trained via scaled RL

While Hugging Face and latent.space detail the technical architecture and training innovations — including Groupwise Agentic Grading, Multi-Prefix Multi-Teacher On-Policy Distillation, and the release of ~7,000 RL environments with automatic graders — only the-decoder.com reports Anthropic’s accusation that Xiaomi improperly used Claude model interactions to distill training data, citing over 400,000 exchanges in March–April 2026 as part of a broader claim involving seven Chinese labs.

All sources agree on the model’s specifications, open release of tools and training tasks, and its position as a leading open-weight model scoring 46 on Artificial Analysis’s Intelligence Index with strong cost efficiency at $0.435/M input and $0.87/M output tokens.