Glonce

Xiaomi's MiMo-V2.6 tops open models as Anthropic alleges Claude misuse

modelsThe Glonce editorsThe Decoder🔥74Glonce rating74Impact70Novelty58Relevance95Surprise63Players78Credibility85
Xiaomi's MiMo-V2.6 tops open models as Anthropic alleges Claude misuse

Xiaomi released its MiMo-V2.6 lineup, led by MiMo-V2.6-Pro, which now tops the rankings of openly available AI models on Artificial Analysis's Intelligence Index, scoring 46 points and landing ahead of rivals such as Kimi K3 and Qwen. The model is also far cheaper than similarly capable competitors: $0.435 per million input tokens and $0.87 per million output tokens, which by Artificial Analysis's math puts a single test task at about $0.13. Artificial Analysis places MiMo-V2.6-Pro on what it calls the Pareto frontier of intelligence and cost. Pro is a mixture-of-experts model with 1.02 trillion total parameters, of which only 42 billion are active per request; a smaller, more efficient MiMo-V2.6-Flash sits alongside it.

Xiaomi attributes the jump to a heavily expanded round of reinforcement learning (RL), training through trial, feedback and reward, scaled along three axes: more data per training step, more varied task environments, and more compute spent grading the solutions. The run took less than six days and cost about $2.62 million for Pro and $0.85 million for Flash, according to Xiaomi. On the DeepSWE coding benchmark, Pro's score rose from 58.4 to 72.6 and Flash's from 48.8 to 65.7. To keep training stable at this scale, Xiaomi froze the model's internal distribution mechanism and added several layers of protection against reward hacking, the tricks a model uses to game its reward signal without actually solving the task.

Alongside the two models, Xiaomi released an especially fast variant, Pro-UltraSpeed, with up to 20 times the output speed, and opened up its RL toolkit: the technical report, the full training framework, a smaller model meant for further training, and about 7,000 ready-made training tasks with automatic graders covering software development, cybersecurity, office work and web design, plus roughly 1,000 tasks for music composition. The tasks draw on a mix of sources: some code comes from real GitHub pull requests by Xiaomi employees and from user queries, other task descriptions are generated by a language model, the cybersecurity tasks draw on OSS-Fuzz (a collection of tens of thousands of real software vulnerabilities), and the office environments are rebuilt synthetically.

This show of openness sits in sharp contrast to accusations Anthropic raised about two weeks before this article. In a threat intelligence report covering cases of Claude abuse discovered between December 2025 and August 2026, Anthropic named seven Chinese labs tied to campaigns against the model: Alibaba, Moonshot AI, DeepSeek, Zhipu, Xiaomi, MiniMax and SenseTime. Across all seven, Anthropic says the labs generated about 190 million exchanges to siphon off Claude's capabilities for training their own models, a technique the company calls illegal distillation.

Xiaomi is named specifically. In a case tagged GTG-16008, Anthropic tracked more than 400,000 exchanges over 20 days in March and April 2026, in which Xiaomi routed user conversations and coding sessions from its own MiMo models through OpenClaw and OpenCode into Claude, aiming to enrich training data for future models. The report offers little documentation of where the earlier training and teacher data for Xiaomi's internal distillation of teacher models came from. Put together, the report's claim is that Xiaomi recorded user conversations with its MiMo models and fed them into Claude to extract training data, the same data that has now helped push MiMo-V2.6 to the top of the open-model rankings.

Key facts

  • MiMo-V2.6-Pro scores 46 on Artificial Analysis's Intelligence Index, the top mark among open models, at $0.435 per million input tokens and $0.87 per million output tokens, about $0.13 per test task.
  • Pro is a 1.02 trillion-parameter mixture-of-experts model with 42 billion active parameters per request; training took under six days and cost $2.62 million for Pro and $0.85 million for Flash.
  • Xiaomi open-sourced its RL toolkit, including about 7,000 automatically graded training tasks and roughly 1,000 more for music composition.
  • Anthropic says seven Chinese labs, Xiaomi among them, generated about 190 million exchanges to extract Claude's capabilities for training, a technique it calls illegal distillation.
  • In case GTG-16008, Anthropic tracked more than 400,000 Xiaomi exchanges routed through OpenClaw and OpenCode into Claude over 20 days in March and April 2026.

Why it matters

MiMo-V2.6-Pro's showing puts an openly available model on the Pareto frontier of cost and intelligence, ahead of other open leaders like Kimi K3 and Qwen, at a fraction of what similarly capable models charge per task. Xiaomi frames the gain as a straightforward payoff from scaled reinforcement learning: more training data per step, more varied task environments, and more grading compute, run in under six days. Anthropic's report complicates that framing. If part of the training signal behind MiMo-V2.6 came from routing MiMo conversations through Claude, a practice Anthropic calls illegal distillation, then the model's price and performance gains are not purely the product of Xiaomi's own RL recipe.

Who it affects

Xiaomi, whose lineup gains a genuine open-model leadership claim and a fresh accusation to answer, alongside six other Chinese labs Anthropic names in the same report: Alibaba, Moonshot AI, DeepSeek, Zhipu, MiniMax and SenseTime. Anthropic, whose Claude the report says Xiaomi used to extract training data for its own models. Developers evaluating open models on cost and capability, and anyone building on the RL toolkit and training tasks Xiaomi has released publicly.

How to use it

MiMo-V2.6-Pro is priced at $0.435 per million input tokens and $0.87 per million output tokens, with MiMo-V2.6-Flash offered as the cheaper, smaller option, and a Pro-UltraSpeed variant available for up to 20 times the output speed. Xiaomi has released the RL toolkit itself: the technical report, the full training framework, a smaller model meant for further training, and around 7,000 automatically graded tasks spanning software development, cybersecurity, office work and web design, plus roughly 1,000 for music composition, all available to build on directly.

How solid is it

The performance and pricing figures come from Artificial Analysis's independent Intelligence Index and cost calculations, not from Xiaomi's own claims alone, which gives the open-model ranking some outside verification. The training cost, duration and DeepSWE benchmark improvements are Xiaomi's own reported numbers. The distillation accusation comes from Anthropic's threat intelligence report, which names a specific case (GTG-16008) with an exchange count and a date range, but the article notes the report offers little documentation of where the earlier training and teacher data for Xiaomi's internal distillation came from, and it includes no response from Xiaomi.

Risks and caveats

The article gives no publication date for either Xiaomi's release or Anthropic's report, only relative timing ('two weeks earlier'). No Xiaomi rebuttal is included, so the distillation accusation stands unanswered here. No figures are given for MiMo-V2.6's market adoption or revenue, only benchmark and cost numbers. The 190 million exchanges figure covers all seven named labs combined, not Xiaomi alone; Xiaomi's own tracked figure is the separate 400,000-plus exchanges in case GTG-16008.