Xiaomi's MiMo-V2.6-Pro tops the open-weights index, but its own table puts Opus 5 ahead
Xiaomi trained it with one reinforcement learning run across coding, agents, vision and cybersecurity, and put the weights on Hugging Face under MIT.
On September 22, 2026, Xiaomi listed three new models on its own release page, and the weights for two of them are on Hugging Face. The license on the repo is MIT, so you can run the model commercially, change it and ship it without asking Xiaomi.
The flagship is MiMo-V2.6-Pro-RL. It's a mixture of experts model with 1.02 trillion parameters in total and 42 billion working on any one token, and it reads video and audio as well as text and pictures.
- Series
- MiMo-V2.6, from Xiaomi
- License
- MIT
- Architecture
- Sparse MoE, 1.02T total / 42B activated parameters
- Context Length
- 1M tokens
- Modalities
- Text, Image, Video, Audio
- Third model
- mimo-v2.6-pro-ultraspeed, up to 20x faster
What Xiaomi says it did differently
The pitch on the model card is about training rather than size. Coding, general agents, vision and cybersecurity all went into one reinforcement learning run instead of a run each, and Xiaomi says the tasks were mixed into the same batch so that the skills would reinforce each other.
The other half of it is grading. A pass-or-fail score can't rank two attempts that both pass, so Xiaomi put an agentic grader in the loop: it compares each attempt against the others in its group and builds a rubric for the task out of the contrast. (The card's names for those two pieces are Groupwise Reward Synthesis and Groupwise Advantage Redistribution.)
scaling RL compute, environment diversity, and grader compute together, so the model keeps expanding its capability frontier through exploration and feedback
Where an outside index put it
Artificial Analysis had the model through its own benchmark within hours of the weights going up, and published where it landed and what a task on it costs.
MiMo-V2.6-Pro debuts as the top open weights model on the Artificial Analysis Intelligence Index (46). At $0.13 per Intelligence Index task, it lands on the Intelligence vs. Cost per Task Pareto frontier @Xiaomi has just released MiMo-V2.6-Pro, an open weights model with major advances in intelligence over its predecessor, MiMo-V2.5-Pro (Intelligence Index: 26). Despite the improvement, it retains the same attractive pricing at $0.435 per 1M input tokens (with a 99% cache-hit discount) and $0.87 per 1M output tokens. This makes MiMo-V2.6-Pro one of the most cost-efficient models to deploy. MiMo-V2.6-Pro is an MoE model with 1.02T total parameters and 42B active parameters. Stay tuned for additional analysis of the model. Check out MiMo-V2.6-Pro full benchmarking breakdown here: https://t.co/czBJhQKuWJ
The table Xiaomi printed
The model card carries a comparison against Claude Opus 5, GPT-5.6 Sol and Claude Fable 5. Opus 5 has a number on thirteen of those rows. It's ahead on nine of them, MiMo-V2.6 Pro takes three, and the two tie at 31.6 on Agents' Last Exam.
Where MiMo wins it wins narrowly: 53.1 to 50.3 on AutomationBench v1.0.6, and 89.9 to 89.1 on Terminal Bench 2.1. The losses aren't all narrow, though. Terminal Bench 4.0 goes 34.9 against 49.0, and ExploitBench 47.9 against 70.0.
The cybersecurity block surprised me. Xiaomi describes the flagship as built for cybersecurity among other things, and it scores 94.0 on CyberGym and 80.2 on Xiaomi's own MiMo Cyber Bench. The cheaper Flash model beats it on the first of those, 95.1 to 94.0. Neither row carries a number for Opus 5 or GPT-5.6 Sol, so the table gives you nothing to set those two scores against.