Unverified50% confidenceFactExact time
DeepSeek V4.1 uses local attention for some layers, looking only at the most recent 128 Tokens, and short-term notes are not archived but kept only in local memory and cleared after about three minutes, with the hit rate still above 98%.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/21/2026
First Seen
Valid until: 1/19/2027
Sources
DeepSeek Open-Sources Inference Framework DeepInfer: The Secret to Halving AI Inference Costs
bilibilicode秘密花园10/21/2026
Related Entities
Related Claims
UnverifiedIn everyday use of DeepSeek Harness, the KV cache hit rate for consecutive tasks is basically above 90%, and a second question on the same 60,000-plus-token document waits only 0.4 seconds.74% similarUnverifiedTesting shows that DeepSeek V4.1 Flash 1017 reduced its thinking token count by about one-third, producing more concise output.72% similarUnverifiedDeepSeek V4.1 Flash reduces per-token storage from 3,500 bytes in the previous generation to approximately 890 bytes through KV cache inter-layer alternation and reuse optimization.71% similarUnverifiedLlama 3.1 70B uses 8 KV heads instead of 64 attention heads, reducing the KV cache size to about one-eighth of the original.69% similarUnverifiedDeepSeek 1017's expert weights are already FP4 (4 bits per parameter), so conventional 4-bit quantization is of little help; the Engram memory table can be placed on SSD (reading only about 12KB per token), but the expert weights must reside in high-speed memory, because the router selects 6 of 384 experts per layer and makes a different selection for each token.68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1081776API
curl https://kongchang.com/api/v1/knowledge/claims/1081776MCP
get_claim(id=1081776)