Xiaomi has released MiMo-V2.6, a series of omnimodal models under an MIT license that includes the flagship MiMo-V2.6-Pro. The Pro model utilizes a sparse mixture-of-experts (MoE) architecture comprising 1.02 trillion total parameters, with 42 billion active parameters per token distributed across 384 routed experts (eight active per token). The technical backbone features 70 layers blending global and sliding-window attention, supplemented by dedicated vision and audio encoders. Capable of processing text, image, video, and audio with a one-million-token context window, the Pro model achieved a 46.32 on the Artificial Analysis Intelligence Index v4.3, surpassing open-weights competitors like GLM-5.3 and Kimi K3. To ensure reproducibility, Xiaomi open-sourced the training environments, RL code, and technical reports, having live-streamed a six-day reinforcement learning run. During this training, the models completed 30 RL steps across 750,000 trajectories, utilizing an asynchronous architecture that processed up to 3.7 billion tokens per step. This process yielded significant improvements on the DeepSWE v1.1 benchmark, where Pro rose from 58.4 to 72.57. The release includes the lighter MiMo-V2.6-Flash and the high-speed MiMo-V2.6-Pro-UltraSpeed, aiming to challenge closed-source systems by providing high-performance, self-hostable intelligence at competitive API pricing.