Xiaomi's MIT-licensed MiMo-V2.6-Pro-RL tops open-weights intelligence index, discloses ~$2.6M RL training cost
lmoroney · x · 2026-10-03
Laurence Moroney highlights Xiaomi's newly shipped open-weights lineup, all under the MIT license: MiMo-V2.6-Pro-RL tops Artificial Analysis' open-weights Intelligence Index, alongside Flash and a 9B distill.
Key training details:
- In coding RL tasks, once a change passed its tests, an AI grader also scored how focused the diff was and whether it matched the repo's conventions—rewarding code like a careful reviewer would.
- Xiaomi published 7,000+ RL task environments plus the training code.
- The Pro RL run cost roughly $2.6M.
Moroney's takeaway: if you teach or build agents, put reward design on the syllabus—measure whether changes are small, focused, and consistent with the codebase, the way a careful human reviewer would.
Related event: Xiaomi's MiMo-V2.6 Tops Open-Source Model Rankings with $2.6M Training Cost(2 posts)→
More from coding & agent
- Devs flag GPT's defensive programming for hiding bugs; OpenAI asks for /feedback cases — pvncher · 2026-11-19
- Flutter ships A2UI 1.0: agents respond with real interactive widgets, not chat text — rseroter · 2026-11-19
- Skills vs MCP vs RAG vs Memory: the four context mechanisms AI agents actually need — MaryamMiradi · 2026-11-19
- Anthropic details agent memory: 'dreaming' lets Claude organize its experience — daniel_mac8 · 2026-11-19
- The case for externalized memory as the real fix for continual learning — daniel_mac8 · 2026-11-19
- FT: US tech outages blamed on rush to fix bugs found by AI — nordicinst · 2026-11-19