Xiaomi's MiMo-V2.6 Pro Is Now the Top Open-Weight Model

Xiaomi has released MiMo-V2.6, a pair of open-weight models that take the top spot among downloadable models on the Artificial Analysis Intelligence Index. The flagship, MiMo-V2.6 Pro, scores 46 — the same as Grok 4.7, a closed model released the same day. Its smaller sibling, MiMo-V2.6 Flash, shipped alongside it. Both weights went up on Hugging Face on September 21 under an MIT licence, which puts no restriction on commercial use.

Pro is a sparse mixture-of-experts model with 1.02 trillion total parameters, 42 billion of them active per token, routed across 384 experts with eight firing at a time. Flash is 309 billion total and 15 billion active. Both read and write across a one-million-token context and are natively omnimodal: text, images, video and audio go into the same model, through a 681-million-parameter vision encoder and a separate 308-million-parameter audio tokenizer, rather than being bolted on as adapters. The backbone mixes sliding-window and global attention — 60 of Pro’s 70 layers use a 128-token local window — and a five-layer multi-token-prediction head does speculative decoding.

Xiaomi’s headline claim is that Pro performs "on par with Claude Opus 5 and GPT-5.6 Sol across most agent benchmarks." The company’s own table mostly bears that out. It beats both on AutomationBench (53.1 against 50.3 and 45.8), ties Opus 5 on Agents’ Last Exam at 31.6, and edges it on Terminal Bench 2.1, 89.9 to 89.1. On DeepSWE v1.1, the coding-agent test most labs now lead with, it lands at 71.9, two points behind Opus 5 and one behind Sol.

The weak spots are just as clear in the same table. Terminal Bench 4.0, the harder successor, has Pro at 34.9 against Opus 5’s 49.0. And on the offensive-security benchmarks — writing working exploits rather than finding bugs — the gap opens to 20 or 30 points. That is a capability labs have increasingly chosen to gate in their own closed models, so an open model trailing there is not obviously something Xiaomi will be rushing to close.

The more consequential number may be the price. Through Xiaomi’s API and OpenRouter, Pro costs $0.435 per million input tokens and $0.87 per million output; Flash is $0.14 and $0.28. Artificial Analysis puts Pro’s cost to run its full index at about $0.13 per task. Xiaomi also offers a Pro-UltraSpeed tier at ten times the price, which it says generates up to 20 times faster at the same quality — a speed play it first showed in June, when it pushed a trillion-parameter model past 1,000 tokens a second.

According to VentureBeat, the reinforcement-learning stage that turned the base models into agents cost about $2.62 million for Pro and $850,000 for Flash, running roughly 750,000 trajectories in under six days. Xiaomi has also published the RL framework and the harnesses behind it — composable environments for coding, office workflows, visual work and cybersecurity — and a 9-billion-parameter distilled model built on Qwen. Fuli Luo, who leads the MiMo team after leaving DeepSeek, framed the release around scarcity: "In an era when compute is brutally scarce, we still chose to dedicate a team of several dozen people to one goal."

There is one caveat worth holding onto. Every benchmark figure above is Xiaomi’s, and the closed models in its comparison are already a generation old: Anthropic shipped Claude Opus 5.5 and OpenAI shipped GPT-6 Sol the day after MiMo-V2.6 went up. So "on par with the frontier" means on par with where the frontier stood last week. That is still a new position for an open model — and at roughly a fifth of GPT-6 Sol’s input price, it is one that anyone can download, fine-tune and run on their own hardware.