Unverified70% confidenceFactExact time
DeepSeek-V2 proposes MLA (Multi-head Latent Attention), which compresses KV Cache to approximately 5% of MHA through low-rank compression.
2
Sources
70%
Confidence
Long-term
Relevance
7/15/2026
First Seen
Sources
Related Claims
UnverifiedDeepSeek's innovations in Multi-head Latent Attention (MLA) and FP8 mixed-precision training compress KV Cache memory usage to 5%–13% of standard multi-head attention.79% similarVerifiedDeepSeek V4 adopts a proprietary hierarchical compressed attention strategy, featuring two alternating attention modes: CSA (Compressed Sink Attention) and HCA (Heavy Compressed Attention).79% similarUnverifiedDeepSeek-V2引入Multi-Head Latent Attention（MLA）机制和DeepSeekMoE架构，其每百万token推理成本约为同期GPT-4级别模型的百分之一74% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/512131API
curl https://kongchang.com/api/v1/knowledge/claims/512131MCP
get_claim(id=512131)