MiMo V2.5 Inference Optimization

The MiMo-V2.5 model family adopts an architecture combining Hybrid SWA (Hybrid Sliding Window Attention), MoE (Mixture of Experts), and a multimodal encoder. This secures both performance and efficiency simultaneously in long-context scenarios.

Hybrid SWA alternates SWA layers, which use a local window, with Full Attention layers, which see the entire context. In the case of MiMo-V2.5-Pro, 60 of the total 70 layers are configured as SWA, reducing Compute and KVCache storage space to about 1/7 compared to Full Attention.

To implement these theoretical advantages in an actual production environment, the following engineering optimizations were carried out.

• KVCache Management: Applied a hierarchical caching system optimized for the SWA structure and a Prefix Cache Tree
• Scheduling: Optimized the Prefill/Decode execution pipeline and synchronized distributed cache state
• Multimodal Optimization: Resolved throughput bottlenecks that occur when processing large-scale images and long videos

As a result, the MiMo-V2.5 series successfully implements the efficiency of its theoretical architecture in an actual inference system, demonstrating excellent performance in long-context reasoning and multimodal understanding capabilities.