Xiaomi Open-Sources MiMo-V2.6, a 1.02-Trillion-Parameter Model That Tops the Open-Weights Leaderboard

Xiaomi has officially released and open-sourced its MiMo-V2.6 series, the smartphone giant's most aggressive push yet into frontier-scale AI. The family consists of two natively fully multimodal models, Pro and Flash, with weights published on Hugging Face under an MIT license that permits commercial use — a notable stance from a company whose core business gives it plenty of reasons to keep model cards close.

The flagship checkpoint is a sparse mixture-of-experts model containing 1.02 trillion total parameters, with 42 billion activated per token over a 1-million-token context window. On the Artificial Analysis Intelligence Index, MiMo-V2.6-Pro scored 46, placing it first among open-weights systems, ahead of Z AI's GLM-5.3 at 45 and tying the proprietary Grok 4.7. A gap to the strongest closed models remains — Claude Fable 5.1 and GPT-6 Astra still lead the index — but the ordering at the open frontier has clearly changed. Xiaomi says its Pro model outperformed Kimi K3 and Qwen3.8 Max to take the open-source crown.

The release is framed around recursive self-improvement: rather than treating RL as a final polish, Xiaomi scaled reinforcement-learning compute across verifiable complex tasks and let the models improve through iterative feedback loops. The company documented the process in public, live-streaming a six-day RL training run in which both models completed 30 steps each over roughly 750,000 cumulative trajectories, at training costs of around $850,000 for Flash and $2.62 million for Pro.

On agentic benchmarks, Xiaomi claims MiMo-V2.6-Pro performs on par with Claude Opus 5 and GPT-5.6 Sol on most agent evaluations. Published scores include 89.9 on Terminal Bench 2.1, 82.0 on OSWorld-Verified and a striking 94.0 on CyberGym, with the Flash variant even edging Pro on that security benchmark at 95.1. The company also says its thinking-mode chains of reasoning shrank 13 to 30 percent on math benchmarks, cutting token cost without sacrificing accuracy.

Pricing is unchanged from the V2.5 generation: MiMo-V2.6-Pro costs $0.435 per million cache-miss input tokens and $0.87 per million output tokens, while Flash runs $0.14 and $0.28 respectively. Cache writes are free for a limited time, and a Pro UltraSpeed mode offers up to 20x output speed at premium rates. Xiaomi claims that at the same intelligence level, its pricing is 1/20th to 1/60th of comparable overseas models — a new cost-performance record, by its measure, for a domestic large language model.

Both models are now generally available across AI Studio, MiMo Desktop and the company's API platform, with RMB pricing listed separately for China. Artificial Analysis calculated the Pro model's operating cost at $0.13 per index task, noting that running the full evaluation cost $206.66.

For developers, the practical takeaway is leverage: an MIT-licensed, trillion-parameter model that matches proprietary systems on agent benchmarks gives enterprises negotiating room against closed APIs, and gives startups a frontier-class foundation they can self-host. The pressure now sits on the labs charging premium rates — if a phone maker can open-source this, the question is what the next open checkpoint looks like.