Xiaomi releases MiMo-V2.6-Pro and Flash MoE models with Agentic RL report

On October 8, the MiMo team at Xiaomi released a technical report titled "MiMo-V2.6: Scaling Reinforcement Learning Towards Self-Improvement," describing a large-scale Agentic RL training process. The approach scales the training batch, task environments and harness, along with the Grader Compute used to judge agent behavior, in order to explore a path toward continuous model self-improvement.

The release covers two models, MiMo-V2.6-Pro and MiMo-V2.6-Flash. Pro has 1.02T total parameters and activates 42B parameters per inference; Flash has 310B total parameters and activates 15B. Both use a sparse Mixture-of-Experts (MoE) architecture and mix local sliding window attention (SWA) with global attention (GA) in the text backbone, supporting training on context lengths in the million-token range.

According to the report, MiMo-V2.6 splits reinforcement learning compute into three parts: Rollout, Grading and Training, each scaled independently. In practice, every RL step first samples 1568 prompts, and each prompt generates 16 rollouts at once, producing roughly 25,000 trajectories per step and 2.7 billion to 3.7 billion tokens in total. The RL post-training compute cost is about $2.6 million.

In an October 11 report, InfoQ Chinese described the series as the first to systematically scale Agentic RL simultaneously across the three dimensions of Rollout, environment and Grader, and noted that the million-token context is used to support long-horizon agent behavior modeling. The report also links to the technical report at https://arxiv.org/pdf/2610.11959. Public information currently stops at the technical report and the two model releases; no later versions or deployment plans were mentioned.