Xiaomi Publishes HySparse 2 Core and Says MiMo‑V3 Will Adopt It

Overview

• Xiaomi's MiMo team released the HySparse 2 architecture and announced MiMo‑V3 will use it, with technical details published on arXiv for community review.
• Xiaomi reports HySparse 2 cuts prefill FLOPs to about one fifth and shrinks persistent KV cache to roughly one quarter when working at roughly one million tokens of context.
• The design adds KV Bridging and KV Reuse plus token‑level selection and a forced recent‑token window so decoder and encoder K/V states can be shared and prefill can stop earlier.
• Xiaomi says these changes raise retrieval scores (MRCRv2 and RULER‑v2) and lower AgentPPL and LongPPL, which it ties to better performance for interactive agent workloads that issue short actions but receive long observations.
• Alongside HySparse 2, Xiaomi open‑sourced MiMo‑V2.6 (Pro and Flash) as part of an effort to scale reinforcement‑learning compute for iterative model improvement and to lower latency and memory costs for long‑context applications.