MiMo-V2.6-Pro-RL: 1.02T total / 42B active (Sparse MoE); MiMo-V2.6-Flash-RL: 309B total / 15B active (Sparse MoE); MiMo-V2.6-Distill-Qwen-9B: 9B
1M tokens (Pro-RL and Flash-RL; Distill-Qwen-9B not separately specified)
MIT
Official open-weight MiMo-V2.6 family release led by two native omnimodal reinforcement-learning checkpoints, MiMo-V2.6-Pro-RL and MiMo-V2.6-Flash-RL, plus the compact MiMo-V2.6-Distill-Qwen-9B agent model. Pro-RL and Flash-RL jointly process text, images, video, and audio and support a 1M-token context window. Pro-RL uses a 1.02T-parameter sparse MoE backbone with 42B active parameters, while Flash-RL uses 309B total parameters with 15B active; both incorporate Xiaomi's MiMo ViT, separate audio tokenization and patch encoders, hybrid sliding-window and global attention, and a five-layer Multi-Token Prediction speculative decoder. Xiaomi describes the training recipe as 'You Only RL Once': one mixed reinforcement-learning run spanning coding, general agent tasks, visual interaction, and cybersecurity, with tasks and multiple agent harnesses combined within the same batches. Its fully asynchronous GRPO system processed 1,568 prompts with 16 rollouts each and approximately 3.5–3.7B tokens per training step. The release introduces Groupwise Reward Synthesis, Groupwise Advantage Redistribution, and Multi-Prefix Multi-Teacher On-Policy Distillation, while freezing the MoE router and applying adversarial reward-hacking defenses. Across six days, Flash and Pro each completed 30 RL steps, producing approximately 750,000 combined trajectories at reported training costs of $850,000 and $2.62 million respectively. Xiaomi reports that live RL raised DeepSWE v1.1 from 48.8 to 65.7 for Flash and from 58.4 to 72.6 for Pro; final model-card results include 67.9 and 71.9 on DeepSWE, 87.6 and 89.9 on Terminal-Bench 2.1, and 80.8 and 82.0 on OSWorld. Xiaomi additionally reports an Artificial Analysis Intelligence Index score of 46 for Pro and describes it as the strongest open-weight model at release. MiMo-V2.6-Distill-Qwen-9B is a supervised fine-tune of Qwen3.5-9B on 77.4B MiMo-generated tokens, including 27.2B loss-bearing tokens, targeting coding, general computer-use agents, visual coding, and cybersecurity.
https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-RL/resolve/main/MiMo_V2_6_technical_report.pdf