Xiaomi Details HySparse2 Sparse Attention for MiMo-V3
Xiaomi Details HySparse2 for MiMo-V3: Two-Level KV Sharing Cuts 1M-Token Prefill FLOPs About 5× vs Hybrid SWA
Two-level KV sharing targets the long-context prefill bottleneck that agentic workloads hit, cutting compute without a full attention redesign. Engineers building long-context inference stacks can benchmark the approach against hybrid SWA baselines on similar MoE setups. The MiMo-V3 linkage also signals how Xiaomi positions its model stack against domestic rivals like DeepSeek and Qwen.
Xiaomi's LLM-Core team has published details on HySparse2, a hybrid sparse-attention architecture built around two-level KV sharing.
The design targets MiMo-V3-class agentic models and was evaluated on an 80B-A3B MoE configuration.Xiaomi reports roughly 5× lower FLOPs for 1M-token prefill compared with Hybrid SWA.
Xiaomi Details HySparse2 for MiMo-V3: Two-Level KV Sharing Cuts 1M-Token Prefill FLOPs About 5× vs Hybrid SWA
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.
This report is categorized under AITOP's "Papers" desk, focusing on China's AI industry and domestic foundation model ecosystem. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows and US-China foundation model benchmarks and commercialization comparison. Key entities and technical topics under continuous tracking include Xiaomi, MiMo-V3, HySparse2, sparse attention.
Primary source verification: published directly by pandaily with first-party authority. Cross-source corroboration: 3 independent sources have covered this development (including IT之家, shao__meng), confirming industry-wide signal strength.
