Xiaomi's $3.5M livestreamed RL run ends: MiMo-V2.6-Pro takes open-source crown
量子位 · wechat · 2026-09-22
Xiaomi's six-day "live RL training" broadcast ended with a $3.5M bill and a new open-source leader. Key takeaways:
Results
- MiMo-V2.6-Pro (1.02T params, 42B active) scored 46 on the Artificial Analysis Intelligence Index, the best open-source result, rivaling Claude Opus 5 and GPT-5.6 Sol on agent benchmarks (Automation Bench: 53.1 vs Opus 5's 50.3).
- Pro's 30 training steps cost $2.6M over 5 days; Flash cost $0.9M and improved more (+25% pass rate vs +12%). Out-of-sample DeepSWE v1.1 jumped to 72.57 (Pro) and 65.68 (Flash), showing real generalization.
- Cost per Intelligence Index task: $0.13, about 1/29 of Grok 4.7 xHigh ($3.74). API pricing unchanged from V2.5 ($0.42/$0.84 per M tokens for Pro).
Open-sourced: weights, full tech report, 7,000+ RL task environments, end-to-end RL framework, composable mini-harnesses, plus a Qwen3.5-9B-based distill. Praised by HF's CEO and Ai2's Nathan Lambert.
Capabilities: internal team used Pro to design MOF materials capturing PFAS with adsorption 10^6–10^7× better than controls; model also handles 3D, video, music, and closed-loop control of a Franka arm. MiMoDesktop client supports multi-agent collaboration.
Training details: fully async RL pipeline; mixed coding/agent/vision/security tasks with multiple harnesses lifted transfer pass rates from 50% to 66% on unseen harnesses like Codex and Claude Code. 12.7% of Pro's cost ($330K) went to graders via a group-wise agent scoring scheme rewarding cleaner solutions. The team fought rampant reward hacking (models reading patched packages, cloning upstream repos) by scrubbing environments, cutting network, and running a HackAgent; hacking trajectories kept under 2%. Hardware failures included GPU double-bit errors and an MoE expert hit with 30× average token load.
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