Xiaomi's MiMo V2.6 Pro Tops Open-Source at $0.13 per Task, One Point Behind GPT-5.6
FellMentKE · x · 2026-09-24
Xiaomi's MiMo V2.6 Pro scored 46 on the Artificial Analysis Intelligence Index — #1 among open-source and Chinese models, ahead of Qwen3.8-Max, Kimi K3 and GLM-5.3, just one point behind GPT-5.6 Sol.
Key points:
- Weighted average cost per index task is $0.13, putting it on the intelligence-cost Pareto frontier; Xiaomi says V2.6 nearly doubles intelligence over V2.5 at unchanged pricing.
- Fully multimodal, built for complex projects, long-horizon tasks, cybersecurity and scientific research.
- Gains attributed to a post-training self-improvement system built around Environment × Task × Grader, pushing the same pre-trained base via RL.
- In practice it has helped materials researchers design and computationally screen candidate materials, alongside coding and multimodal reasoning work.
- The V2.6 family also includes Flash and Pro-UltraSpeed variants.
Related event: Xiaomi's MiMo-V2.6-Pro Tops Open-Source Model Rankings(2 posts)→
More from Models
- GPT-6 Astra Fixes a UI That Even Opus 5.5 Couldn't Crack — haltakov · 2026-09-24
- The coin-flip test: why LLMs fail at calibration, which is the actual product — airesearch12 · 2026-09-24
- JevBench v1.4.1 adds six systems; Jev-Ommi debuts at #7, top five unchanged — airesearch12 · 2026-09-24
- Next-gen Chinese agentic models surface: GLM 5.3, DeepSeek 4.1 Flash, Kimi K3 — zainhas · 2026-09-24
- Perplexity CTO's AutoJev-27B Takes Open-Source Lead on Decision Index, Trained for Just $3k — antoine_chaffin · 2026-09-24
- Leak: early look at 'Grok Bots' tab inside X's web app — nima_owji · 2026-09-24