LLM-Core Xiaomi introduced MiMo-V2.6, a reinforcement learning framework for multimodal agents. It scales RL compute with 1,568 prompts and 2.7–3.7B tokens per step. The system uses Groupwise Advantage Redistribution to rank passing solutions by quality. Freezing the MoE router prevents expert imbalance during training. Both Pro and Flash models improved steadily across coding, workflows, visual tasks and cybersecurity. RL cost $2.6M for Pro and $0.9M for Flash. The grader itself used 12.7% of Pro’s RL compute. A distilled 9B model reproduced the gains on released environments. DeepSWE v1.1 scores rose from 58.4 to 72.6 for Pro and 48.7 to 65.7 for Flash. SWE-bench Verified improved from 61.1 to 66.2 with the distilled model. Cybersecurity mini-benchmark scores rose from 31.3 to 47.0. The paper open-sources training dynamics, RL environments and framework.

• 1,568 prompts per training step
• 16 attempts per prompt
• 2.7–3.7B tokens per step
• $2.6M RL cost for larger model