最新模型
MiMo-V2.6 Releases Open Weights, Uses Group Scoring to Improve Cross-Framework Agent Training
Xiaomi combines multiple types of agent tasks and tool frameworks in a single reinforcement learning run, using scorers to distinguish between solutions that have passed tests. The accompanying 9B distilled model is still an SFT checkpoint, so results from subsequent RL cannot be directly attributed to these weights.
Xiaomi’s September 22 release notes list MiMo-V2.6 Pro and Flash, and the series’ weights are now public. The main technical focus is combining software development, general agent, vision, and cybersecurity tasks in a single mixed reinforcement learning run, while also mixing the tool frameworks used by agents to encourage learned strategies to transfer across frameworks. [Release notes](https://mimo.mi.com/docs/en-US/updates/model), [official announcement](https://mimo.mi.com/docs/en-US/news/latest/v2-6)
The model card describes fully asynchronous GRPO training, with each step containing 1,568 prompts and 16 execution trajectories generated per prompt. The team also increases the compute allocated to scorers: GRS builds task scoring rubrics offline from different trajectories, then combines them with test results; GAR compares solutions that have passed tests online, assigning advantage signals to higher-quality trajectories. This provides a learning signal for shortening action sequences and reducing token usage even after tests have passed. [Training design](https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-RL)
Maintaining training stability also requires additional engineering. Xiaomi says it freezes MoE routers to curb expert load drift and reduces reward hacking through adversarial evaluation, anomaly detection, and verifier cross-checks. Its announcement also presents more than 7,000 agent task environments and end-to-end training resources based on verl and other tools. Whether these materials can fully reproduce the training still requires checking the environments, code, and scoring configurations actually available, item by item. [Official announcement](https://mimo.mi.com/docs/en-US/news/latest/v2-6)
A more accessible starting point for researchers is the simultaneously released MiMo-V2.6-Distill-Qwen-9B. Its model card explicitly identifies it as SFT weights produced by fine-tuning Qwen3.5-9B on MiMo-generated data, intended as a starting point for subsequent agent reinforcement learning. It scores 44.6 on SWE Pro using avg@3 and 37.1 on Terminal Bench 2.1 using avg@1. The subsequent RL results listed in the announcement cannot be treated as the capabilities of these downloadable weights. [9B model card](https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B)
For development teams, this provides a concrete starting point for studying how tasks, agent frameworks, and scorers jointly affect learning. Published scores still come primarily from vendor evaluations, and some test sets are internal. When adopting the models, teams should hold the weights, chat template, tools, and retry budget constant, then examine success rates on Traditional Chinese tasks, total cost per task, and recovery behavior after failures. Comparisons of training methods should also separately record the time spent on environment interactions, scorer calls, and policy updates. Even if fewer tokens are used per task, the added scoring cost may change the overall cost-benefit balance. These are inferences about evaluation design. Claims of self-improvement should likewise be understood as iterative gains constrained by the training environments and reward design.