Xiaomi unveils hysparse2 architecture for enhanced efficiency in long-context agents

On September 24, 2026, Xiaomi announced the release of HySparse2, the core architecture for its MiMo-V3 system, designed to optimize long-context agent performance. The new framework reduces prefill computation and KV cache size while improving accuracy in retrieving key information from growing task histories. HySparse2 builds on its predecessor, HySparse, by introducing two-level KV sharing and token-level sparse selection, enabling more efficient processing of long-context data. The architecture divides the model into two parts: a Self-Decoder with full attention and sliding window attention, and a Cross-Decoder with full attention and sparse attention. This allows for shared KV caches and reduces the need for separate local branches. In tests on an 80B-A3B MoE model, HySparse2 reduced prefill computation by 1/5 compared to Hybrid SWA and by 1/3 compared to HySparse, while also cutting KV cache size from 12GB to 2.7GB. The improvements enhance long-context retrieval and reduce agent prediction errors, marking a significant step forward in efficient multi-turn reasoning.