By utilizing the MIT License, Xiaomi mentioned it’s permitting industrial deployment, continued coaching, and fine-tuning with out further authorization. Tulika Sheel, senior vp at Kadence Worldwide, mentioned the MIT License could make it enticing. “It permits enterprises to freely modify, deploy, and commercialize the mannequin with out restrictions, which is uncommon in in the present day’s AI panorama,” Sheel mentioned.
“On ClawEval, V2.5-Professional lands at 64% Go^3 utilizing solely ~70K tokens per trajectory — roughly 40–60% fewer tokens than Claude Opus 4.6, Gemini 3.1 Professional, and GPT-5.4 at comparable functionality ranges,” Xiaomi mentioned in a weblog publish.
The fashions use a sparse mixture-of-experts (MoE) design to handle compute prices. The 310-billion-parameter MiMo-V2.5 prompts solely 15 billion parameters per request, whereas the 1.02-trillion-parameter Professional model prompts 42 billion. Xiaomi mentioned the Professional mannequin’s hybrid consideration design can scale back KV-cache storage by almost seven instances throughout long-context duties.