
Key points
Xiaomi MiMo HySparse2 attention and cache reuse | Prefill runs 25 layers and decode 49; each cross FA selects 1,024 global plus 128 recent tokens
Image source: @SemiAnalysis_

Xiaomi MiMo HySparse2 attention and cache reuse | Prefill runs 25 layers and decode 49; each cross FA selects 1,024 global plus 128 recent tokens
Image source: @SemiAnalysis_