MiMo-V2.6-Pro-RL is the flagship checkpoint of Xiaomi's MiMo-V2.6 series, a sparse Mixture-of-Experts model built on 384 routed experts with 8 activated per token for 1.02T total and 42B activated parameters. It is natively omnimodal, taking text, image, video, and audio in one model, and supports a 1M-token context window intended for long repositories, tool traces, and multi-session agent runs. The backbone has 70 transformer layers (60 sliding-window attention, 10 global attention) at a 6144 hidden size, paired with a 681M-parameter MiMo ViT vision encoder and an audio stack of a 308M-parameter AudioTokenizer encoder with 20 RVQ codebooks plus a separate 127M-parameter audio patch encoder; a 5-layer sliding-window multi-token-prediction (MTP) module speculatively drafts 7 tokens per forward pass. Xiaomi calls its training recipe "You Only RL Once": one mixed reinforcement-learning run spanning coding, general-agent, visual, and cybersecurity tasks, using fully asynchronous Group Relative Policy Optimization (GRPO) at 1,568 prompts x 16 rollouts per step. To rank passing solutions rather than score them pass/fail, Xiaomi adds Groupwise Agentic Grading: Groupwise Reward Synthesis builds task rubrics offline from contrasting rollouts, and Groupwise Advantage Redistribution ranks passing trajectories online against the policy's own samples, followed by a Multi-Prefix Multi-Teacher On-Policy Distillation stage extending coverage to harder-to-verify tasks. On Xiaomi's own published benchmark table, MiMo-V2.6-Pro-RL scores 89.9 on Terminal-Bench 2.1 and 31.6 on Agents' Last Exam, both the top or tied-top figures among the listed comparison models (Claude Opus 5, GPT-5.6 Sol, Claude Fable 5), and 53.1 on AutomationBench v1.0.6, ahead of all three; it also rises from its MiMo-V2.5-Pro predecessor's 19.0 to 71.9 on DeepSWE v1.1. On several other benchmarks it trails those same competitors, including OSWorld-Verified (82.0 vs up to 86.0) and ExploitBench (47.9 vs up to 78.5). Xiaomi ships serving recipes for SGLang and vLLM, and the README's reference SGLang command spans two nodes. The model card lists the MIT license.