MiMo-V2-Flash is a Mixture-of-Experts language model developed by Xiaomi, combining 309 billion total parameters with a sparse activation design that puts roughly 15 billion parameters to work during any given forward pass. This architectural choice allows the model to deliver substantial capability while keeping inference costs manageable. The design centers on a hybrid attention approach that interleaves sliding window attention with global attention in a 5-to-1 ratio, using a 128-token sliding window to achieve near-six-fold reduction in KV-cache storage compared to naive global attention. Multi-Token Prediction extends the model's ability to generate multiple tokens per step, accelerating throughput for high-speed reasoning and agentic workflows. The model also offers a hybrid-thinking toggle, letting users control reasoning behavior through a simple boolean flag.

MiMo-V2-Flash represents Xiaomi's open-source foundation model lineage, released under an open-weight framework that has attracted community tooling including vLLM integration recipes and pipeline configurations. On software engineering benchmarks—specifically SWE-bench Verified and SWE-bench Multilingual—the model ranks as the top-performing open-source option globally, delivering performance that rivals Claude Sonnet 4.5 while operating at roughly 3.5% of the cost. A 256K context window supports extended reasoning chains and multi-step agent tasks. The design philosophy prioritizes practical agent scenarios, coding tasks, and complex reasoning chains, positioning the model as a versatile backbone for both research and production applications requiring open-weight access and cost-efficient deployment.