Unverified50% confidenceBenchmarkExact time
从V1到V4.1 Flash，单Token缓存从389000字节压缩到890字节，总共压缩了437倍
1
Sources
50%
Confidence
Long-term
Relevance
9/11/2026
First Seen
Sources
Related Entities
Related Claims
Unverified经过压缩后，DeepSeek V4 Flash模型体积从接近500GB降至80.8GB73% similarUnverifiedV4 Flash的缓存大小只有V3.2的7%71% similarUnverified以quality=75压缩一张569,826字节的JPEG图片后，文件缩小到376,630字节，压缩率为33%71% similarUnverifiedDeepSeek V4 Flash的UD-Q8_K_XL量化版本文件体积为162GB，采用MXFP4+BF16混合精度，达到完全无损71% similarVerified即便用 4-bit 量化把权重压缩到 214GB，仍远超单张 48GB 显卡的容量70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/900713API
curl https://kongchang.com/api/v1/knowledge/claims/900713MCP
get_claim(id=900713)