MiMo-V3 adopts HySparse2 for cheaper long context
MiMo-V3 is getting a new architecture: core HySparse2 cuts prefill and KV cache and improves long-context retrieval versus MiMo-V2.6 Hybrid SWA, with 5.02× lower prefill FLOPs at 1M tokens.
KV cache
All AI Pulse updates tagged "KV cache".
1 signals