Xiaomi reveals HySparse2, a two-level KV sharing design for MiMo-V3-class models
According to Pandaily, Xiaomi's LLM-Core team published HySparse2, a hybrid sparse-attention design with two-level KV sharing aimed at MiMo-V3-class agentic models, reporting about 5× lower 1M-token prefill FLOPs versus Hybrid SWA on an 80B-A3B MoE. The article notes the HuSparse2 design targets efficiency in large models, though details on deployment are not provided.
Key facts
- 01Xiaomi's LLM-Core team published HySparse2.
- 02HySparse2 is a hybrid sparse-attention design with two-level KV sharing.
- 03It targets MiMo-V3-class agentic models.
- 04It reports about 5× lower 1M-token prefill FLOPs versus Hybrid SWA on an 80B-A3B MoE.
AI-generated from the sources below. Always check the originals.
How each country tells it
So far our sources show coverage from one country. Perspectives appear once we find reports from a second country.
Sources
Summaries are AI-generated from the linked sources and may contain errors; always check the originals. We summarise and link; we never republish articles. Photos come from openly licensed libraries, official publicity material and brand logos, credited to their sources. If you own an image and want it credited differently or removed, email info@coda.news and we will act promptly.