MiMo-V2-Flash

MiMo-V2-Flash is built on a sparse Mixture-of-Experts architecture that utilizes 309 billion total parameters, with 15 billion active parameters per inference. To balance computational efficiency with high-level performance, the model employs a hybrid attention mechanism that interleaves a 128-token sliding window with full attention in a 5:1 ratio. This design is specifically engineered to support advanced reasoning, software engineering, and tool invocation, positioning the model as a foundational technology for complex agentic tasks and interconnected device ecosystems.

The development of this model prioritizes robust agent execution and collaborative reasoning, supported by stable reinforcement learning techniques during post-training. By focusing on these areas, the model achieves top-tier results on benchmarks such as the AIME 2025 math competition and the GPQA-Diamond scientific knowledge test, while also leading in software engineering evaluations like SWE-bench. These capabilities reflect a strategic shift toward models that can interact effectively with physical and digital environments, making it a strong candidate for developers building sophisticated, multi-step autonomous systems.