Xiaomi MiMo-V2.6 Pro and Flash Under the MIT License

MiMo-V2.6-Pro has 1.02 trillion total and 42 billion active parameters. Xiaomi published RL code; Artificial Analysis scored Pro 46 on its Intelligence Index.

A phone maker's trillion-parameter open model

Xiaomi dated its MiMo-V2.6 announcement September 22, 2026. It covers two natively omnimodal models, MiMo-V2.6-Pro and MiMo-V2.6-Flash, plus an UltraSpeed mode for Pro that Xiaomi says runs up to 20 times faster. Pro's model card lists a sparse mixture-of-experts design with 1.02 trillion total and 42 billion active parameters, text, image, video and audio input, and a 1M-token context window, and the weights are released under the MIT license. Latent Space, publishing on September 22, embedded Artificial Analysis results posted the previous evening. [1][4][2]

Prices are unchanged from MiMo-V2.5: $0.435 per million input tokens and $0.87 per million output tokens for Pro, and $0.14 and $0.28 for Flash. The models launched in Xiaomi's AI Studio, MiMo Code and MiMo Desktop, which left early access with this release, and through Xiaomi's API and OpenRouter. [1]

The measurement and the recipe

Artificial Analysis scored Pro 46 on its Intelligence Index, the highest of any open-weights model it had measured and up from 26 for MiMo-V2.5-Pro. Its leaderboard puts Pro a fraction of a point behind Grok 4.7 at its xhigh setting, a proprietary model released on September 21. [3][2]

Xiaomi released its technical report, training environments and reinforcement learning code with the models. It says the RL run for each model took under six days, 30 steps over about 750,000 trajectories, at a cost of about $0.85 million for Flash and $2.62 million for Pro, and lifted Pro's score on DeepSWE v1.1 from 58.4 to 72.57. Latent Space noted that the complete sets of more than 7,000 training tasks had not yet been published. [1][2]

Atlas interpretation: The $2.62 million is the cost of the final reinforcement learning run, not of the model; pretraining is not in it, so it is not the price of a frontier-class model. The published environments and code are the more durable contribution, since they let other labs test the recipe. The release came twelve days after Anthropic's misuse report said Xiaomi had replayed user conversations with its own MiMo models to Claude. [1][5]

Sources

  1. Introducing MiMo-V2.6 series

    Xiaomi · Sep 22, 2026

  2. [AINews] Xiaomi MiMo-V2.6-Pro 1T-A42B: the new top Open Weights model, trained for $3M

    Latent Space · Sep 22, 2026

  3. MiMo-V2.6-Pro - Intelligence, Performance & Price Analysis | Artificial Analysis

    Artificial Analysis

  4. XiaomiMiMo/MiMo-V2.6-Pro-RL

    Hugging Face · Sep 22, 2026

  5. Detecting and countering misuse of AI: September 2026

    Anthropic · Sep 10, 2026