Xiaomi opens MiMo-V2.6 live RL training panel: $1.16M+ spent, rewards still climbing
APPSO · wechat · 2026-09-17
- Xiaomi MiMo lead Luo Fuli published a live RL training dashboard for MiMo-V2.6, a rare look inside the training black box. Pro and Flash have collectively cost over $1.16M ($30.9K/hour), consuming 25.1B and 40.5B tokens respectively.
- Three scaling directions: 2B tokens per step with 1568 prompts × 16 rollouts in fully async training; multi-task agentic RL mixing visual, code, general and chat environments; agentic in-group credit assignment with test-case and rubric-based rewards.
- Panel shows mean rewards climbing (Pro 0.55→0.582, Flash 0.52→0.570), DeepSWE v1.1 scores of 63.72 (Pro) and 60.77 (Flash), and average context length nearing 100K tokens. Training continues, with more technical details promised in 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