MiMo-V3 introduces HySparse2, a new architecture optimizing agentic inference through two-level KV sharing, achieving 5× lower prefill FLOPs, 4.5× smaller KV cache, and improved long-context retrieval compared to MiMo-V2.6.