Xiaomi's MiMo-V2.6 report details $2.6M single-run RL post-training
Xiaomi's MiMo team published a technical report, "MiMo-V2.6: Scaling Reinforcement Learning Towards Self-Improvement," covering two sparse MoE models: MiMo-V2.6-Pro (1.02T total parameters, 42B activated) and MiMo-V2.6-Flash (310B total, 15B activated). The report says each RL step samples 1,568 prompts and roughly 25,000 agent trajectories totalling 2.7–3.7B tokens, with RL post-training costs of about $2.6M for Pro and $900K for Flash. The team also open-sourced more than 7,000 RL environments plus an end-to-end RL framework.
Xiaomi scales RL across environments, graders and agent harnesses rather than just model size, and publishes cost and failure details that are rarely disclosed, making the recipe easier for others to reproduce.
10月8日，小米 MiMo 团队发布技术报告《MiMo-V2.6: Scaling Reinforcement Learning Towards Self-Improvement》。这篇报告披露了小米一套完整的规模化 Agentic RL 训练过程：通过同时扩大训练 Batch、任务环境与 Harness，以及用于评判 Agent 行为的 Grader Compute，来探索模型持续自我改进的路径。
此次发布包含 MiMo-V2.6-Pro 和 MiMo-V2.6-Flash 两个模型。其中，Pro 总参数量为 1.02T、每次推理激活 42B 参数；Flash 总参数量为 310B、激活 15B 参数。两者均采用稀疏混合专家（MoE）架构，并在文本主干中混合局部滑动窗口注意力（SWA）与全局注意力（GA）。
论文链接：
https://arxiv.org/pdf/2610.11959"
同时 Scale 了 Rollout、环境和 Grader
MiMo-V2.6 将强化学习算力拆成三个部分：Rollout、Grading 和 Training。
#Xiaomi#Xiaomi MiMo#MiMo-V2.6-Pro#MiMo-V2.6-Flash#Hugging Face#verl#mini-swe-agent#Codex#Claude Code#DeepSWE#Artificial Analysis