MiMo-V2-Flash: High-Efficiency Inference, Code and Agent Foundation Model (platform.xiaomimimo.com)

Xiaomi has announced the open-sourcing of MiMo-V2-Flash, a cutting-edge MoE (Mixture of Experts) model designed for exceptional inference efficiency. With 309 billion total parameters and 15 billion activated parameters, this model integrates a hybrid attention mechanism and a multi-layer MTP (Multi-Token Prediction) architecture. It performs at the top tier of open-source models across various agent evaluation benchmarks, showcasing coding capabilities that rival Claude 4.5 Sonnet while achieving a mere 2.5% inference cost compared to it. The model dramatically accelerates generation speed, making it a significant advancement for the AI/ML community. The MiMo-V2-Flash's architectural innovations, such as hybrid attention combining Global Attention and Sliding Window Attention (SWA), allow effective management of large context lengths, with performance improvements in general tasks and reasoning. Furthermore, the MTP mechanism enables parallel validation of tokens, leading to real-world speedups of 2.5 to 3.7 times. The fully open-sourced model weights and inference code under the MIT license, along with a free API for a limited time, are poised to encourage widespread experimentation and deployment, pushing the boundaries of cost-effective AI solutions.