Luo Fuli, who leads Xiaomi’s MiMo large model development team, has announced that MiMo-V3 will feature a cutting-edge new architecture, centered around the official release of its core component, HySparse 2. This innovative architecture minimizes the need for pre-filling, reduces the size of KV caches, and enhances long-context retrieval capabilities significantly. When handling a 1M Token length, it cuts pre-fill computation by 5.02 times and shrinks the KV cache by 4.5 times. Concurrently, there are improvements in MRCRv2 and RULER-v2 scores, along with reductions in AgentPPL and LongPPL. The HySparse 2 architecture is tailored specifically to the demands of agent inference workloads, integrating key mechanisms such as KV bridging and KV reuse. Furthermore, the architecture introduces two notable enhancements: transitioning from block-level selection to Token-level selection, and replacing independent SWA branches with a forced window for recent Tokens. These refinements enable both local and global Tokens to utilize the same KV cache, and allow pre-filling to cease once completed within the self-decoder. Previously, Xiaomi released and open-sourced the MiMo-V2.6 series of native full-modal models, marking a significant milestone in its pursuit of recursive self-improvement pathways.