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
- User automates entire video clipping workflow with Nous Research's Hermes agent — Teknium · 2026-11-12
- As models improve, writing good evals gets harder and more important — amankhan · 2026-11-12
- Prime agent ships with day-one multi-agent communication, runs in Minecraft unmodified — sarahcat21 · 2026-11-12
- Autonomous agent edits just 1,024 weights to lift Qwen3.5 0.8B from 7.4% to 37.4% on MATH-500 — VictorSanh · 2026-11-12
- Omarchy 4.6 lets you share local AI models across your tailnet, says DHH — alexcovo_eth · 2026-11-12
- Coding harnesses hand shell access to inference providers — 4-step gateway mitigation — shensi · 2026-11-12