Xiaomi's Luo Fuli Recaps MiMo-V2.6 RL Run: 25K Agent Trajectories Per Step
bookwormengr · x · 2026-09-22
Xiaomi's Luo Fuli published a long retrospective on MiMo-V2.6, framed around one question: pretraining scaling laws have run for years — can RL scale the same way, with more compute, more tasks and more real environments still translating into capability gains? Xiaomi has spent the past half year betting on it.
RL scale
- Each update samples 1,568 tasks, with 16 rollouts per task — roughly 25,000 agent trajectories and billions of tokens per step.
- Both Pro and Flash ran 30 steps.
Challenge 1: enough "worlds"
Math RL is one problem, one answer, instant feedback. Agents must write code, open browsers, operate computers, call tools and read results before deciding the next move. Xiaomi expanded its environments accordingly, building 7,000+ RL task environments spanning coding, general agent, visual agent and cybersecurity tasks, with trajectories lasting dozens to hundreds of steps.
Challenge 2: Harness enters training
Previously the harness lived outside the model — model makers trained the model, and products like Claude Code and Codex wrapped a system prompt around it. Xiaomi brought the harness into the training loop.
(The original post is a long thread; later sections were not fully provided.)
More from coding & agent
- Pydantic Ships Monty V1 Beta: Major Wire Protocol Refactor Ahead of 1.0 — samuelcolvin · 2026-09-22
- Stale Data Was an Analytics Problem; Agents Turn It into an Operations Problem — bigdata · 2026-09-22
- Astra builds a full stage lighting rig in Sentinel, outputting real lighting data — PurzBeats · 2026-09-22
- Talk Preview: Post-Train AI Models on Your Own Domain with RL Environments — ben_burtenshaw · 2026-09-22
- His AI-Built Website Confused Users, So He Redesigned the UI by Hand in 4 Days — maddy30445r · 2026-09-22
- I Built a Techmeme for AI Agents — Agents Read, Post and Rank It, Humans Only Watch — akash_kloudle · 2026-09-22