MiMo-V2-Flash is a Mixture-of-Experts foundation language model designed to deliver strong reasoning, coding, and agentic performance while remaining efficient enough for practical deployment. It uses a total of 309B parameters with 15B active per inference, paired with a hybrid attention architecture that interleaves sliding-window and full attention at a 5:1 ratio using an aggressive 128-token sliding window. This design aims to preserve long-context capability while keeping compute and memory requirements relatively low, and the model is fully open-sourced, allowing the community to inspect, run, and build on the weights directly.

In benchmark evaluations, MiMo-V2-Flash ranks among the top two open-source models on the math competition AIME 2025 and the scientific knowledge benchmark GPQA-Diamond, signaling strong general reasoning ability. On SWE-bench Verified and Multilingual, it claims the number-one spot among open-source models for software engineering tasks and performs competitively against leading frontier systems. Its successor, MiMo-V2.5, is built directly on the MiMo-V2-Flash backbone, extending it with dedicated vision and audio encoders into a native omnimodal model that inherits the same hybrid attention scheme, suggesting the architecture was designed to scale cleanly into future multimodal and agentic applications.