Xiaomi Open-Sources MiMo V2.6: Top Open-Source Model Matches Grok 4.7 at 1/30th the Cost
新智元 · wechat · 2026-09-23
Xiaomi's MiMo team released and open-sourced the MiMo V2.6 series (Pro/Flash/Ultraspeed). Headline numbers: Pro scores 46 on the AA Intelligence Index — matching Grok 4.7 at $0.13 vs $3.74 per task (30x cheaper), claiming the top open-source spot globally. Lead researcher Luo Fuli calls it likely the largest single RL training run ever open-sourced, with challenges 'exceeding DeepSeek R1.'
The three models: Pro is a flagship omni-modal model beating Opus 5 and GPT-5.6 Sol in several Design Arena categories; Flash surpasses the previous V2.5 Pro flagship; Ultraspeed keeps full Pro performance with 500–1000 TPS.
Real-world results: Blender 3D modeling, interactive open-world generation, deck creation, and a materials-science case where Pro designed MOF structures capturing PFAS at up to millions of times reference performance, cutting a month-long cycle to 2–3 days.
MixRL training: The team live-streamed RL training (750K trajectories, $850K/$2.62M costs; DeepSWE: Flash 48.8→65.68, Pro 58.4→72.57). Three axes: async rollouts (1568 prompts ×16, 2.7–3.7B tokens/step, 4000 GPUs), mixed environments (68% coding across 25 sources, 21 harnesses), and an Agentic Grader doing in-group ranking and credit assignment that suppresses reward hacking. Prices stay at V2.5 levels, redrawing the intelligence-cost Pareto frontier.
More from Models
- YC-backed Monte launches post-training platform to build proprietary models from your agents' real work — ycombinator · 2026-11-12
- 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