Xiaomi's MiMo-V2.6 lands as open MIT weights, omnimodal and agentic

A 309B/15B-active MoE that near-ties Opus 5 on agentic coding, a trillion-param Pro sibling, and a 9B distill you can actually self-host tonight.

Nowline SEP 22 3:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • Flash: 309B sparse, 15B live, near-frontier on agents

    MiMo-V2.6-Flash is a 256-expert MoE (8 active, ~15B per token) with a 1M-token context and native text/image/video/audio. It scores 67.9 on DeepSWE v1.1 and 80.8 on OSWorld-Verified — within a few points of Opus 5 — and leads on CyberGym at 95.1. The FP8 weights are 172.9 GB across 65 shards.

  • Pro is the trillion-param sibling, same MIT license

    MiMo-V2.6-Pro-RL dropped 18 seconds after Flash under the same MIT terms. Vendor evals put it near ~72.6 on DeepSWE — a statistical tie with Opus 5, GPT-6 Astra, and Gemini 3.8 Flash. You'll need cluster-scale GPUs, but the weights are yours to fine-tune and self-host with no revenue cap.

  • Build this weekend: the 9B distill runs on one GPU

    MiMo-V2.6-Distill-Qwen-9B fine-tunes Qwen3.5-9B on MiMo traces, pushing SWE-Pro from 32 to 44.6 and AutomationBench from 5 to 30.3. It's the version most builders can actually serve — a cheap, self-hosted agent brain for coding and tool-use with no API bill.

  • The catch: no price list, no GGUF yet

    Xiaomi still hasn't published API pricing, so any hosted MiMo-V2.6 endpoint number you see is someone's guess. There are no GGUF quants either — the full models want SGLang or vLLM in FP8, and the 9B distill is BF16-only for now, so llama.cpp users are still waiting.

  • Elsewhere: ChatGPT's new cookie tracks you across the web

    Security researchers report an OpenAI '__obi' ad cookie that follows ChatGPT accounts onto third-party advertiser sites, backed by concrete request evidence. If you embed ChatGPT or ship a browser agent, cross-site tracking is now in scope for your privacy review.