Unverified50% confidenceFactExact time
Only four source layers (kv_source_layer_ids = [2,8,14,20]) provide long-range KV, and a compression ratio is applied at each layer.
1
Sources
50%
Confidence
Long-term
Relevance
10/18/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedQuantizing the KV cache from FP16 to Q4 can achieve approximately 4x VRAM savings at the cost of minimal precision loss.71% similarUnverifiedModels using MLA such as DeepSeek-V4.1-Flash compress the KV state to about 2.4 GB per 32k stream (FP8 KV), while wider architectures consume 3.0 GB–3.8 GB per 32k stream.67% similarVerifiedDeepSeek-V4.1-Flash's KV cache is compressed to 890 bytes per token, only one-quarter that of V4 Flash.66% similarUnverifiedQ4_K_M量化文件大小17.77GB，pp2048速度96.8 tok/s，tg128速度11.4 tok/s64% similarUnverifiedAgens Volundr 32B is a dense architecture with approximately 32B parameters and 72 layers in total, of which only 18 layers retain KV cache.62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1061810API
curl https://kongchang.com/api/v1/knowledge/claims/1061810MCP
get_claim(id=1061810)