Xiaomi released MiMo-V2.6-Pro-RL. It is a sparse mixture-of-experts model with 1.02T total parameters. It handles text, image, video, and audio in one model. It supports a 1M token context window. It uses reinforcement learning to scale self-improvement across coding, agents, visual, and cybersecurity tasks. It employs Groupwise Reward Synthesis and Groupwise Advantage Redistribution for grading. It includes aligned RL with self-correction and adversarial screening. It is available on Hugging Face and ModelScope. It can be deployed via SGLang or vLLM. It scores 71.9% on DeepSWE v1.1 and 89.9% on Terminal Bench 2.1.
What it unlocks
Running a single model across text, image, video, and audio with long-horizon reasoning and agentic tasks.
- 1.02T total parameters, 42B activated
- 1M token context length
- 71.9% on DeepSWE v1.1
- 89.9% on Terminal Bench 2.1
What you need to act on it
- self-hosting or API access
- SGLang or vLLM deployment
Sources