XiaomiMiMo/MiMo-V2-Flash

MiMo-V2-Flash is a Mixture-of-Experts language model built to balance massive scale with efficient execution. It features 309 billion total parameters while utilizing only 15 billion active parameters per forward pass, a design choice aimed at delivering high-speed reasoning. The architecture incorporates a hybrid attention mechanism that interleaves sliding window attention with global attention, which helps reduce the KV cache footprint significantly. By integrating Multi-Token Prediction, the model achieves faster inference speeds, making it well-suited for agentic workflows that require both depth and responsiveness.

The model was pre-trained on 27 trillion tokens and utilizes a Multi-Teacher On-Policy Distillation paradigm to refine its capabilities. This post-training method allows the model to learn from domain-specialized teachers, effectively absorbing expert knowledge to improve performance on complex tasks. With its ability to handle extended context lengths, the model is positioned to compete with frontier-level systems in coding and reasoning benchmarks. Its open-source nature and technical design make it a versatile tool for developers looking to integrate high-capacity, efficient AI into their own infrastructure.