Xiaomi releases top-ranked open AI model as Anthropic accuses it of Claude distillation
MiMo-V2.6-Pro beats competitors on performance and price, but Anthropic says Xiaomi used 400,000+ Claude queries to train it.
What to know
- Xiaomi's MiMo-V2.6-Pro ranks as the strongest openly available AI model while undercutting competitors on price ($0.13 per task vs. fractions of dollars for equivalently capable models).
- Anthropic accused Xiaomi of illegal distillation, documenting 400,000+ exchanges in March–April 2026 where Xiaomi routed MiMo outputs through Claude to extract training signal.
- The performance gains came from $2.62 million in reinforcement learning, completed in under six days, scaled across data, task variety, and compute for grading.
- Xiaomi is opening its reinforcement learning toolkit, including 7,000 training tasks and the full training framework, in stark contrast to the secrecy Anthropic alleges.
“All told, the labs are said to have generated about 190 million exchanges to siphon off Claude's capabilities for training their own models, a technique Anthropic calls illegal distillation.”
The Decoder · The Decoder ↗ · Sep 21
Xiaomi AI model developerAnthropic Claude AI developerArtificial Analysis AI benchmarking firm
How it unfolded 1 development · click the chart to see its coverage articles
-
background
Xiaomi open-sources its reinforcement learning toolkit and training tasks — Alongside the models, Xiaomi released an especially fast Pro-UltraSpeed variant with up to 20 times the output speed, plus its full RL training framework, technical report, a smaller model for further training, and about 7,000 ready-made training tasks with automatic graders for software development, cybersecurity, office work, web design, and roughly 1,000 tasks for music composition.
-
1
Xiaomi releases MiMo-V2.6 lineup topping open-source benchmarks
Xiaomi released MiMo-V2.6-Pro and MiMo-V2.6-Flash, with Pro scoring 46 points on Artificial Analysis's Intelligence Index and ranking as the strongest openly available AI model. Pro costs $0.435 per million input tokens and $0.87 per million output tokens, undercutting similarly capable competitors. The 1.02 trillion parameter mixture-of-experts model was trained with $2.62 million in reinforcement learning over less than six days.
“The tricks a model uses to game rewards without actually solving the task.”
— The Decoder · source -
1 outlet Xiaomi's affordable flagship AI leads the open models, and Anthropic says Claude helped get it there
first by The Decoder, 4d ago
-
-
background
Anthropic publishes threat intelligence report naming Xiaomi in distillation campaign — Anthropic released a threat intelligence report examining Claude abuse between December 2025 and August 2026, naming seven Chinese labs—Alibaba, Moonshot AI, DeepSeek, Zhipu, Xiaomi, MiniMax, and SenseTime—as having generated about 190 million exchanges total to siphon off Claude's capabilities using what Anthropic calls illegal distillation.
-
background
Xiaomi routes MiMo outputs and conversations through Claude — Over 20 days in March and April 2026, Xiaomi passed user conversations and coding sessions from its own MiMo models through OpenClaw and OpenCode to Claude (case GTG-16008), generating more than 400,000 exchanges.
-
background
Anthropic detects Claude abuse across seven Chinese AI labs — Anthropic begins tracking cases of Claude abuse, including campaigns by Xiaomi, that would be documented in a threat intelligence report spanning December 2025 to August 2026.