Xiaomi Previews HySparse2 for MiMo-V3: KV Cache Cut to ~1/4.5 at 1M Context

智东西 · wechat · 2026-09-24

Xiaomi MiMo lead Luo Fuli pre-announced HySparse2, a core architectural component of the upcoming MiMo-V3, with a paper published just two days after the MiMo-V2.6 release.

HySparse2 targets three pain points of long-horizon agent inference: heavy Prefill compute, large KV cache, and long-context retrieval. Its two-level KV sharing (KV Bridging + KV Reuse) lets Prefill exit after running only the first half of the model — a 49-layer model needs just 25 layers on prefill nodes, nearly halving weights — while switching from block-level to token-level sparse selection for precise cross-context retrieval.

At 1M-token context, versus HybridSWA, Prefill compute drops to 1/5 and KV cache to 1/4.5 (2.69GB vs 12.09GB); RULER-v2 improves by 18.65 points after post-training, with lower AgentPPL and LongPPL. General capabilities remain roughly on par.

Related event: Xiaomi Unveils HySparse2 Architecture for MiMo-V3, Slashing KV Cache(4 posts)→

Original post →

More from Infra

Infra channel →