Xiaomi Open-Sources MiMo-V2.6 After a $3.47M RL Run, Claims Top Open-Model Score
APPSO · wechat · 2026-09-22
Xiaomi released and open-sourced MiMo-V2.6 (Pro and Flash). Pro scores 46.32 on Artificial Analysis Intelligence Index v4.3, the highest for an open model, beating Kimi K3 and Qwen 3.8 Max, though it trails closed frontier models on TerminalBench and security evals. Xiaomi published the full training: 30 RL steps per model, $3.47M total cost (Pro $2.62M, Flash $850K), 2.5K long sequences per step with up to 1M-token contexts. Research lead Luo Fuli called the engineering harder than DeepSeek R1. The release includes 7,000 RL task environments, verifiers, an end-to-end RL framework, and a distilled Qwen3.5-9B variant. Pricing: Flash output ¥2/M tokens, Pro ¥6, with an UltraSpeed tier at 10x cost for up to 20x speed. Demos include 3D world-building, robot arm control, and Lean4 theorem formalization.
More from coding & agent
- As models improve, writing good evals gets harder and more important — amankhan · 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
- Slack's David Hoang: agents beat people on coordination, proposes 'boundary agents' — davidhoang · 2026-11-12
- CS student mourns coding flow states he'll never experience in the AI era — chrissyykat · 2026-11-12
- AI Engineer Code Summit features its first all-female keynote lineup — altryne · 2026-11-12