MiMo-V2.6-Pro

Xiaomi's flagship omnimodal reasoning model, released as open weights under MIT on September 21, 2026 after a publicly livestreamed RL run. It debuted as the top open-weights model on the Artificial Analysis Intelligence Index at 46 (v4.3), level with Grok 4.7 at the rounded score and a hair behind on the raw index. Same backbone family as MiMo-V2.5-Pro: 1.02T total / 42B active sparse MoE, 70 layers (60 sliding-window with a 128-token window, 10 global), hidden size 6144, 384 routed experts with 8 active and no shared experts, a 5-layer DFlash-style MTP drafter that proposes 7 tokens per pass, a 681M MiMo ViT and 308M + 127M audio encoders. Text, image, video and audio input; 1M context, 128K max output. Pretrained on 30T tokens (27T text, 3T omni) with AdamW, context extended 32K to 256K mid-run, MXFP4 quantization-aware training during mid-training; the released checkpoint is FP8.

The technical report, "Scaling Reinforcement Learning Towards Self-Improvement," is the substance: $2.6M of RL post-training on thousands of GPUs in a single mixed run ("You Only RL Once") across coding, general-agent, visual, and cybersecurity environments under multiple harnesses, versus $0.9M for the Flash sibling. Fully asynchronous GRPO at 1,568 prompts × 16 rollouts per step (2.7–3.7B training tokens per update, contexts to 1M), a frozen MoE router for stability, and groupwise agentic grading: Groupwise Reward Synthesis builds rubrics offline from contrasting rollouts and Groupwise Advantage Redistribution re-ranks passing trajectories online, so the reward scales with grader compute (12.7% of spend; rollout 43.8%, training 43.5%). A multi-prefix multi-teacher on-policy distillation stage (MOPD2) follows. Xiaomi open-sources the RL environments, the RL framework, and the training dynamics, and ships MiMo-V2.6-Distill-Qwen-9B (an SFT of Qwen3.5-9B on MiMo data) as a research starting point.

Reported scores: DeepSWE v1.1 71.9 (Claude Opus 5: 74.0, Fable 5: 70.0), Terminal-Bench 4.0 34.9, Terminal-Bench 2.1 89.9, OSWorld-Verified 82.0, GDPval-AA 2.1 Elo 1673, Toolathlon-Verified 76.9, AutomationBench 53.1, CyberGym 94.0, SEC Bench Pro 66.3. API pricing is $0.435/M input and $0.87/M output (cache hits $0.0036), with the same checkpoint also sold as MiMo-V2.6-Pro-UltraSpeed at up to 20× output speed for $4.35/$8.70. Companion MiMo-V2.6-Flash (309B / 15B active, 48T pretraining tokens) lists at $0.14/$0.28. Led by Luo Fuli's LLM-Core team.