MiMo-V2.6-Flash-RL: Scaling Reinforcement Learning Toward Self-Improvement

MiMo-V2.6-Flash-RL is the efficiency-balanced checkpoint of Xiaomi’s MiMo-V2.6 series. It is a sparse Mixture-of-Experts model with 309B total parameters and 15B active per token. Text, images, video, and audio go through one model, and a 1M token context covers long repositories, tool traces, and multi-session agent runs.

One RL run. Coding, general agents, visual tasks, and cybersecurity train together in a single mixed reinforcement learning run instead of separate per-domain runs. Fully asynchronous Group Relative Policy Optimization (GRPO) runs on large batches: 1,568 prompts with 16 rollouts per step. Mixing tasks and harnesses in the same batch lets capabilities reinforce each other and transfer to harnesses the model never saw in training.

Scaling the reward signal. Pass/fail grading cannot rank two solutions that both pass. Groupwise Reward Synthesis (GRS) builds task-specific rubrics offline from contrasting rollouts and combines rubric quality with test outcomes. Groupwise Advantage Redistribution (GAR) ranks passing trajectories online and moves advantage toward higher-quality solutions, which steers the model toward shorter paths and fewer tokens per task.

Architecture. The backbone has 48 Transformer layers, 39 with sliding-window attention and 9 with global attention, over 256 routed experts with 8 active per token. Vision runs through a 681M parameter MiMo ViT. Audio uses a 308M AudioTokenizer and a 127M audio patch encoder. A 5-layer speculative decoder predicts 7 tokens per forward pass for parallel verification. After mixed RL, Multi-Prefix Multi-Teacher On-Policy Distillation (MOPD2) extends capabilities to hard-to-verify tasks.

Reference

MiMo-V2.6 Technical Report