Xiaomi MiMo-V2.6-Pro tops open-weight models at 46, and it's MIT

Two MIT models — a 1T-param Pro tied with Grok 4.7 and a leaner 310B Flash — ship with 1M context, sub-$1 API pricing, and open weights you can self-host.

Nowline SEP 23 8:00 PM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • #1 open weight, tied with Grok 4.7

    MiMo-V2.6-Pro scores 46 on the Artificial Analysis Intelligence Index — the top open-weight model, level with Grok 4.7 and well ahead of DeepSeek V4.1 Pro's 36. It's a 1.02T-parameter MoE with 42B active, under an MIT license you can actually ship on.

  • Pro or Flash, pick your budget

    Two sizes launched together: the 1T/42B Pro and a leaner 310B/15B Flash. Both are MIT-licensed with a 1M-token context, text/image/audio/video input, and up to 128K output tokens — the cheap one isn't a stripped-down toy.

  • Sub-$1 to call, free to self-host

    Xiaomi's API runs Pro at $0.435/$0.87 and Flash at $0.14/$0.28 per million tokens (in/out). Weights are on Hugging Face, and there's already a distilled MiMo-V2.6-Distill-Qwen-9B for running locally without a datacenter.

  • Frontier open weights for $2.62M

    The Pro came out of one of the largest open RL runs to date — 30 steps over ~750K trajectories in under six days — for a reported $2.62M (Flash: $850K). Training rewarded clean, maintainable solutions over reward-hacking, tuned for long-horizon agent work.

  • The coding numbers

    Where it counts for builders: Terminal Bench 2.1 89.9, DeepSWE v1.1 71.9, AutomationBench 53.1. Flash trails only slightly (DeepSWE 67.9) and posts 95.1 on CyberGym — strong for security-focused agents.