MiMo-V2-Flash is an open-source large language model built by Xiaomi's LLM-Core team as part of the MiMo family, designed around an inference-centric Mixture-of-Experts architecture. The model carries roughly 309 billion total parameters but activates only about 15 billion at any step, an arrangement that prioritizes speed and cost efficiency over dense parameter coverage. Its attention system is hybrid, alternating full attention with sliding-window attention layers in a documented ratio with a 128-token window, which the design notes as helping reduce KV-cache storage substantially while keeping long-range context reachable. Expert routing is handled by a sigmoid-with-bias router with group-limited expert selection and FP32 gate precision, chosen so routing decisions stay stable when the rest of the activations run in lower precision.

For deployment, MiMo-V2-Flash ships as an FP8 Hugging Face checkpoint under the XiaomiMiMo organization, with sensitive components such as layer norms, the gate, the language modeling head, and token embeddings retained in bf16 to preserve quality. The model is positioned for reasoning, coding, and real-time agentic workloads, with reported inference speeds reaching roughly 150 tokens per second and benchmark results cited at 84.9% on MMLU-Pro and 73.4% on SWE-Bench Verified, placing it in competition with leading proprietary systems. A 262,144-token context window supports long-document and multi-step tasks, and the open-weight availability makes the model a practical fit for teams that need a fast MoE backbone they can self-host, fine-tune, or integrate into tool-using pipelines.