MiMo-V2.6 mid-RL run: ~2B tokens per step, multi-harness agentic RL, details to be open-sourced
bodonoghue85 · x · 2026-09-17
Xiaomi's MiMo team (@LuoFuli) says six months of silence went into studying how far RL can scale, and MiMo-V2.6 is now in the middle of its RL run, with training streamed live.
Three things were scaled:
- Compute: 2B tokens per step (1568 prompts × 16 rollouts), fully async
- Environments & harnesses: multi-task agentic RL, mixed across multiple harnesses in one run
- Grader compute: agentic in-group credit assignment with test-case and rubric-based rewards
Details will be open-sourced piece by piece over the coming weeks.
Related event: Xiaomi live-streams MiMo-V2.6 RL training, burning over $1.08M in 36 hours(14 posts)→
More from coding & agent
- Cohere Labs agent team-construction method cuts multi-agent costs by 70% — Cohere_Labs · 2026-11-02
- Long-read: AI's real value comes after the efficiency era—infra glut, agent tooling boom — yangyi · 2026-11-02
- offprompt: MIT-licensed MCP server hands agents API keys without ever exposing them — Vast_Ad_7338 · 2026-11-02
- Reddit idea: expose every OOP public method as AI agent tools to agentify legacy apps — out_of_nowhere__ · 2026-11-02
- DeepSeek design doc: a person can only follow 3-5 agent sessions at once, attention is the real bottleneck — heyshrutimishra · 2026-11-02
- Software fundamentals are key to getting good results from AI coding agents — bendee983 · 2026-11-02