Xiaomi releases MiMo-V2.6-Pro/Flash open models topping open-model rankings, as Anthropic separately accuses Xiaomi of distilling Claude via 400,000+ conversations

Xiaomi released the MiMo-V2.6 model family (Pro, Flash, and a Pro-UltraSpeed variant), with Pro a 1.02T-parameter mixture-of-experts model using 42B active parameters. Xiaomi reports a score of 46 on Artificial Analysis's Intelligence Index, currently the highest among open-weight models, priced at $0.435/M input and $0.87/M output tokens, and says training used expanded reinforcement learning completed in under six days for roughly $2.62M (Pro) and $0.85M (Flash). Xiaomi also open-sourced its RL training framework, a technical report, a smaller model for further training, and around 8,000 auto-graded training tasks. Separately, an Anthropic threat intelligence report published about two weeks earlier alleges Xiaomi routed over 400,000 MiMo user and coding conversations through OpenClaw and OpenCode into Claude during March-April 2026 to harvest training data.

If the benchmark and pricing figures hold up, this gives developers and enterprises a materially cheaper, high-performing open-weight option for coding and general tasks, and the open RL toolkit could lower the barrier for other teams to replicate similar training approaches. But the Anthropic distillation allegation, if substantiated, raises real questions about the legitimacy and reproducibility of Xiaomi's gains and could affect how enterprises and researchers treat MiMo's outputs or trust its benchmark claims. Competing open-model labs (Alibaba, Moonshot, DeepSeek, Zhipu, MiniMax) now face a lower-cost, higher-scoring rival, which matters for anyone choosing between open models on price-performance grounds.

The "best open model" and cost figures are entirely vendor-supplied and unaudited; no independent party has confirmed the training cost, the six-day timeline, or reproduced the benchmark placement. The article's juxtaposition implies the capability leap may be explained by the distillation allegation, but this is speculative framing: the two stories are reported separately by different sources with no established causal link, and the distillation report itself predates this release by two weeks so is not new evidence.

1. 01Whether Artificial Analysis or another independent lab publishes its own MiMo-V2.6-Pro benchmark run, since the 46 Intelligence Index score currently rests on Xiaomi's submission and framing
2. 02Anthropic's next threat intelligence update or any legal/contractual action tied to case GTG-16008, and whether Xiaomi issues a direct response or denial
3. 03Real-world adoption signals: developer uptake of the open RL toolkit and ~7,000 training tasks, and whether other open-model labs (Alibaba/Qwen, Moonshot/Kimi, DeepSeek) match or beat the DeepSWE coding scores within the next few months