Xiaomi's MiMo-V2.6 Pro tops open-source leaderboard at 46 AA, $0.13 per task
SucceededMind · x · 2026-09-24
Xiaomi released and open-sourced the MiMo-V2.6 series (Pro and Flash, both natively fully multimodal). MiMo-V2.6-Pro scored 46 on the Artificial Analysis Intelligence Index, beating Kimi K3 and Qwen3.8 Max to become the strongest open-source model, though still behind closed-source leaders like Claude Fable 5.1 and GPT-6 Astra.
Key points:
- Post-training driven: gains came from a scaled RL self-improvement (RSI) loop — Environment × Task × Grader, 30 RL steps, 750k trajectories, 6 days of live RL training, roughly $2.6M total. Same pretraining foundation, pushed further
- Cost: $0.13 per AA Index task; API pricing unchanged from V2.5, claiming a new cost-performance record among Chinese models
- Packaging: native multimodality plus 1M context
The story: Xiaomi didn't invent new pretraining — it scaled the post-training loop and pushed the intelligence-vs-cost Pareto frontier outward.
More from Infra
- Liquid AI Ships DSpark Draft Model, Speeding Up LFM2.5-VL-3B Decoding Up to 3.13x — JosephJacks_ · 2026-09-25
- DIY Local LLM Cluster: Quad RTX 5070 Ti + Quad 5060 Ti Nodes With vLLM FP8 Inference — Whahine · 2026-09-25
- AMD and Perplexity Partner to Build an 'Agentic PC' on Ryzen AI Halo — AravSrinivas · 2026-09-25
- Water, Noise, Power: What Data Center Critics Get Wrong About Real Costs — bigdata · 2026-09-25
- Powering a home GPU cluster: one PDU safely feeds ~16 RTX 6000 Max-Q cards — TheZachMueller · 2026-09-25
- Docker launches Cloud Sandboxes: microVM isolation for always-on agents, $250 credit — juntao · 2026-09-25