MiMo-V2.6-Pro

MiMo-V2.6-Pro is the flagship omni-modal reasoning model that Xiaomi's MiMo team released and open-sourced in September 2026, the strongest member of the MiMo-V2.6 series. Xiaomi frames the series as a step toward scaling reinforcement learning into self-improvement: RL compute, environment and tool diversity, and grader compute are scaled together in one mixed run across coding, general agents, visual tasks and cybersecurity, rather than separate per-domain runs.

Architecture and specifications

The model is a sparse MoE with 1.02 trillion total parameters and 42B activated per token, 8 of 384 routed experts active; the language backbone has 70 layers (60 sliding-window attention + 10 global attention) with a hidden size of 6144. The models are natively omni-modal: they accept text, image, video and audio input and return text, with a 1M-token context window. The vision encoder is a 681M-parameter MiMo ViT (28 layers, 24 SWA + 4 full attention); the audio stack combines a 308M AudioTokenizer with a 127M audio patch encoder, and a 5-layer MTP speculative decoder predicts 7 tokens per pass.

Published benchmark results

Xiaomi reports the following for MiMo-V2.6-Pro: DeepSWE v1.1 71.9, Toolathlon-Verified 76.9, Terminal Bench 2.1 89.9, Terminal Bench 4.0 34.9, OSWorld-Verified 82.0, JobBench 62.0, Agents' Last Exam 31.6, AutomationBench v1.0.6 53.1, CyberGym 94.0, MiMo Code Bench 63.2, MiMo VisualCoding 72.3, GDPval-AA 2.1 1673. The official table also compares against Claude Opus 5, GPT-5.6 Sol and Claude Fable 5.

Pricing

On the Xiaomi MiMo platform, domestic pricing is ¥3 per million input tokens, ¥0.025 per million cached input tokens and ¥6 per million output tokens; international pricing is $0.435 input, $0.0036 cached input and $0.87 output per million tokens, with a 50% discount for batch inference.

Release and license

The MiMo-V2.6 series was published on Hugging Face on September 21, 2026 under the MIT license, and served through the Xiaomi MiMo platform (mimo.mi.com) from September 22, with distribution also via ModelScope and OpenRouter. Xiaomi published a live dashboard of the RL post-training run: Pro and Flash each took under six days and 30 steps, about 750,000 trajectories in total, at a reported cost of roughly $2.62M and $0.85M respectively.