Xiaomi has released a pair of omnimodal artificial intelligence models designed to push the boundaries of agentic workloads through scaled reinforcement learning. The Pro and Flash versions of MiMo V2.6 now hold the No. 1 and No. 2 positions respectively on the Vals Index for open-weight models. Along with the model weights, the company provided a technical report and access to over 7,000 reinforcement learning task environments to assist the open-source community in agentic research.

The accompanying technical documentation focuses on data and experimental results from the training runs, shifting emphasis away from the architecture-centric reports typically seen from competitors like DeepSeek. Xiaomi live-streamed the training of these models to demonstrate recursive self-improvement in areas such as coding and 3D reasoning. The models have received immediate integration support from vLLM and SGLang for inference.