XNews

MiMo V2.6 Pro Tops Open-Source Leaderboard with RL Training

MiMo V2.6 Pro Tops Open-Source Leaderboard with RL Training

Xiaomi released and open-sourced MiMo-V2.6 Pro and Flash, with Pro scoring 46 on Artificial Analysis intelligence index, tying xAI Grok 4.7 as the strongest open-source model. The model was trained with a large-scale reinforcement learning run (30 steps, 750k trajectories, $2.62M cost) using a MixRL strategy. API pricing is $0.435/M input tokens for Pro and $0.14/M for Flash.

Key insights

MiMo V2.6 Pro Ties Grok 4.7 on Intelligence Index

The Pro version achieved a score of 46 on the Artificial Analysis comprehensive intelligence index, matching xAI Grok 4.7 released the same day. It surpassed Kimi K3 and Qwen3.8 Max, becoming the current strongest open-source model. However, it still lags behind top closed-source models like Claude Fable 5.1 and GPT-6 Astra.

Largest Single RL Training Run for Open-Source Models

MiMo team lead Fuli Luo, who previously worked on DeepSeek R1, stated that V2.6 likely represents the largest single-scale reinforcement learning training by an open-source team. The Pro model underwent 30 major steps covering about 750,000 training trajectories, each averaging 110,000 to 150,000 tokens, with a total training cost of approximately $2.62 million.

MixRL Strategy Combines Diverse Tasks in One Training Run

The team employed a MixRL strategy that mixed coding, vision, and cybersecurity tasks in the same large-scale RL training, rather than training separately and merging. Tasks with difficult verification or long time horizons were trained individually and later combined using the MOPD method.

Aggressive Pricing: 1/20 to 1/60 of Overseas Models

Pro API pricing is $0.435 per million input tokens and $0.87 per million output tokens. Flash is even cheaper at $0.14 input and $0.28 output. Xiaomi claims that at equivalent intelligence levels, prices are 1/20 to 1/60 of overseas models. Both models support a 1 million token context window and multimodal input (text, image, audio, video).

Flash Model Offers High Cost-Effectiveness for Agent Tasks

The Flash model, while significantly cheaper, scores only a few points lower than Pro on many agent benchmarks. For teams needing to run large volumes of AI agent tasks—such as automated code review, document processing, or internal research assistants—Flash may offer better practical value.

Open-Source Release Includes RL Framework and Distilled Model

Beyond model weights, Xiaomi open-sourced the complete RL training framework, over 7,000 RL task environments, and a Qwen-9B distilled model from MiMo RL trajectories. These resources allow other teams to replicate similar RL training on smaller models without training a trillion-parameter base model from scratch.

Source posts