Unverified50% confidenceBenchmarkExact time
DeepSeek在多头潜在注意力（MLA）和FP8混合精度训练方向上的创新将KV Cache内存占用压缩至标准多头注意力的5%-13%
1
Sources
50%
Confidence
Long-term
Relevance
7/24/2026
First Seen
Sources
Related Claims
UnverifiedDeepSeek在推理部署中采用了FP8混合精度量化和Multi-head Latent Attention（MLA）等技术79% similarUnverifiedDeepSeek-V2提出MLA（Multi-head Latent Attention），通过低秩压缩将KV Cache压缩至MHA的5%左右79% similarVerifiedMLA（多头潜在注意力）由DeepSeek团队在DeepSeek-V2中首次提出，通过将Key-Value对压缩到低维潜在空间大幅降低KV Cache显存占用78% similarVerifiedDeepSeek V4采用自研的分层压缩注意力策略，设计了CSA（压缩吸收注意力）和HCA（重度压缩注意力）两种注意力模式交替使用77% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/602698API
curl https://kongchang.com/api/v1/knowledge/claims/602698MCP
get_claim(id=602698)