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.
More from Models
- Zvi dissects Anthropic's latest system card, disputing the autonomy threshold call — Don't Worry About the Vase (Zvi) · 2026-11-12
- antirez backs low-bit DeepSeek 4.1 Flash over Qwen 27B dense on Apple M5 Macs — antirez · 2026-11-12
- Gemini re-renders images as text when vision is removed, dropping only 6% WSR — zacharynado · 2026-11-12
- Bare-harness coding benchmark: Fable 5.2 best overall, Gemini 4 Carbon best value — Semantic_meaning · 2026-11-12
- Google's Physical-Engineering's Last Exam: all 14 frontier models score below 50% — dl_weekly · 2026-11-12
- Reddit's 2026 LLM hall of fame: Opus 5.5 still unbeaten after 9 rival releases — G0dZylla · 2026-11-12