Xiaomi open-sources MiMo-V2.6: biggest RL run yet, priced ~29x below Grok 4.7
智东西 · wechat · 2026-09-22
Xiaomi's LLM team, led by Luo Fuli, released and open-sourced the MiMo-V2.6 series: natively omni-modal Pro and Flash models, with a 20x-faster Pro-UltraSpeed variant to follow, plus open RL environments, training framework, and tech report.
Luo calls it the largest single RL training run for an open model to date, framing it as a step toward RSI on the path to AGI. MiMo-V2.6-Pro scored 46.32 on the Artificial Analysis Intelligence Index—top open model, above Kimi K3 and Qwen3.8 Max, roughly matching Grok 4.7 while costing about 1/29th its price. Training: 30 RL steps each over 6 streamed days, 750k trajectories, $0.85M/$2.62M in compute; innovations include group-based reward scoring, frozen MoE Router, and anti-reward-hacking layers. Capabilities span coding, 3D, Computer Use (robot arm control), design, music, and science tasks like MOF material design and a 6,000-line Lean4 formalization of the Li-Yorke chaos theorem.
Hands-on testing (via OpenCode) found solid basics but slow thinking, redundant file exploration, and occasional errors—overall a value-for-money play.
More from Models
- Leak Claims OpenAI Will Launch Entire GPT-6.2 Family Tomorrow to Counter Opus 5.5 and Fable 5.2 — heypearlai · 2026-11-12
- ARC-AGI-3 leaderboard: models score 90%+ but at $10K+ per task — haider1 · 2026-11-12
- LMArena Launches Agent Arena for Head-to-Head Agent Evaluation — arena · 2026-11-12
- LLMs lack hyperlocal municipal data, explains disputed Texas power outage debate — csuwildcat · 2026-11-12
- Claude API's per-message effort lets you switch reasoning levels without breaking prompt cache — lyraxana · 2026-11-12
- rasbt coins "Benchmaxxing": making the numbers look better (best optics) — infoxiao · 2026-11-12