Unverified50% confidenceBenchmarkExact time
On the Qwen3.5-2B model, DCKV reduces the number of stored bytes per compressed token to more than 55 times lower than bf16, while accuracy loss on three long-context benchmarks does not exceed 1.8 percentage points.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/28/2026
First Seen
Valid until: 1/26/2027
Sources
Related Entities
Related Claims
Unverifiedllama.cpp的--cache-type-k q8_0 --cache-type-v q8_0参数可将KV Cache从FP16量化到8-bit，减少约一半缓存显存占用72% similarUnverifiedChoosing Q4_0 for KV cache quantization can compress the cache size to approximately one-quarter of FP16, with far less impact on output quality than model weight quantization.72% similarUnverifiedvLLM的PagedAttention将KV Cache碎片化率从传统方案的60-80%降至不足4%，吞吐量提升可达24倍71% similarVerified4-bit quantization (such as Q4_K_M) can typically compress the memory footprint of a 3B model from approximately 6GB to approximately 2GB, with a precision loss of less than 2–3%.70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1106217API
curl https://kongchang.com/api/v1/knowledge/claims/1106217MCP
get_claim(id=1106217)