MiMo-V2-Flash is a Mixture-of-Experts language model that combines 309 billion total parameters with a routed architecture, activating only 15 billion parameters per forward pass to keep inference lightweight. The model introduces a hybrid attention architecture that interleaves sliding-window attention with full global attention using a 5-to-1 hybrid ratio and an aggressive 128-token sliding window, enabling efficient long-range reasoning while preserving local context. It incorporates Multi-Token Prediction during pre-training, allowing the model to predict multiple tokens simultaneously and improve overall generation quality. Built specifically for high-speed reasoning and agentic workflows, the architecture prioritizes strong performance on coding and multi-step task execution while remaining computationally efficient.
The model was pre-trained on 27 trillion tokens before entering a novel Multi-Teacher On-Policy Distillation pipeline, where domain-specialized teachers trained via large-scale reinforcement learning provided dense, token-level reward signals to transfer expertise to the student model. This post-training approach enables MiMo-V2-Flash to achieve competitive results despite its modest active parameter count. On SWE-bench Verified and SWE-bench Multilingual, MiMo-V2-Flash ranks as the top open-source model globally, matching the performance of leading closed-source models on software engineering tasks, while also ranking among the top open-source models on math and science benchmarks like AIME 2025 and GPQA-Diamond. Its combination of open weights, efficient architecture, and strong agentic capabilities makes it particularly well-suited for developers building autonomous systems, coding assistants, and reasoning-heavy applications.
