Xiaomi released the MiMo-V2.6 AI model family, and its MiMo-V2.6-Pro scored 83.3 in World of AI Bench, ranking sixth and edging past GPT-5.6 Sol while trailing Claude Opus 5 by about 0.4 points.

The company released two variants, Pro and Flash, and says the family prioritizes reasoning, AI agents, programming, and multi-step tasks. Xiaomi claims MiMo-V2.6-Pro matches models such as Claude Opus 5 and GPT-5.6 Sol on many agent-focused benchmarks, and it has published the model weights.

World of AI Bench is an independent ranking and not a universal industry standard.

Another independent evaluation, Artificial Analysis, scored MiMo-V2.6-Pro at 46.32 and placed it first among open-weight models.

MiMo-V2.6-Pro supports a one million token context window and accepts text, image, video, and audio inputs.

Xiaomi highlighted large-scale reinforcement learning in training. The company said RL training for both models ran for under six days across 30 stages, and that Pro's training cost about $2.62 million. Xiaomi reported that DeepSWE v1.1 rose from 58.4 to 72.6 for the Pro model during the process.

Xiaomi also said API pricing for this generation did not increase versus MiMo-V2.5, and that it has published the weights, a technical report, and RL training resources.