On April 28, 2026, Xiaomi officially released the MiMo-V2.5 series models under the MIT license, enabling their use for commercial purposes, secondary training, and fine-tuning without requiring further authorization. The series entered public beta on April 23 and comprises two models: MiMo-V2.5-Pro and MiMo-V2.5. Both models support context windows on a scale of millions. The MiMo-V2.5-Pro is tailored for complex agent tasks and has claimed the top spot among global open-source models on the GDPVal-AA and ClawEval leaderboards, boasting strong capabilities in code generation and long-range task processing. The MiMo-V2.5 is a native full-modality model that supports text, image, video, and audio understanding, achieving industry-leading task completion rates and token efficiency. Additionally, the series has been optimized for compatibility with chips from multiple manufacturers, including Alibaba T-Head and Baidu Kunlunxin, as well as inference frameworks such as SGLang and vLLM.