Xiaomi launches open-source MiMo-V2.6 Pro & Flash, tops AI index at 46 among open models
burny_tech · x · 2026-09-23
Xiaomi officially introduced MiMo-V2.6, a family of two omnimodal models (Pro & Flash) built through scaled reinforcement learning.
- Benchmarks: Pro reportedly matches Claude Opus 5 and GPT-5.6 Sol across most agent benchmarks, scoring 46 on the Artificial Analysis Intelligence Index — the highest among open-source models
- Improvements: stronger coding, computer use, 3D reasoning, and creative capabilities
- Fully open: weights, technical report, RL environments, and training code are all released
The retweeter adds a notable detail: Xiaomi handles RL reward hacking by red-teaming its own RL environments and iteratively fixing them — better than most approaches, though possibly vulnerable to explicit goal-directed attacks like OpenAI's hacking swarm.
More from coding & agent
- Dev uses Opus 5.5 with Lean to formally verify Claude Agent SDK, yielding 16 bug-fix PRs — bcherny · 2026-09-23
- Opus turns the Lean-based Agent SDK verification workflow into an infographic — bcherny · 2026-09-23
- Chrome flags compromised passwords; user wants a Gemini agent to reset 50+ of them — doooyle · 2026-09-23
- GPT-6 prompt caching gets more reliable: higher hit rates, diagnostics, up to 90% savings — rohanpaul_ai · 2026-09-23
- Grok Bot gets major upgrade: native Google Docs/Sheets/Slides, P90 first-reply 20% faster — XFreeze · 2026-09-23
- Your AI agent ran for 12 hours — but did anyone tell it the brief changed at hour 2? — HaktanSuren · 2026-09-23