Unverified50% confidenceFactExact time
4-bit quantization (such as bitsandbytes NF4 or GPTQ) weights are usually stored in blocks, with each block carrying a scaling factor and zero point, and these metadata are often stored in FP16 or FP32; embedding layers, output layers, and normalization layers may remain at higher precision, so the actual weight file is typically 5%-15% larger than the theoretical P×b/8 value.
1
Sources
50%
Confidence
Long-term
Relevance
10/18/2026
First Seen
Sources
Related Entities
Related Claims
VerifiedThe GGUF format supports multiple precision levels such as 4-bit and 8-bit, and can compress the memory footprint of a 7-billion-parameter model from approximately 14 GB (FP16) to approximately 4 GB (Q4_K_M quantization), with inference accuracy loss typically within 1–3%.77% similarVerifiedQuantized KV caches (such as storing KV vectors in FP8 or INT4) can compress the cache size to 1/2 or even 1/4 of the original.76% similarVerifiedQ4_K_M uses mixed precision, storing some critical weights at 6-bit and the remainder at 4-bit, with an overall average of approximately 4.5-bit.75% similarUnverified4F² is a measure of area efficiency for DRAM memory chips, meaning each memory cell occupies only 4 times the minimum feature size in area, representing the theoretical limit of storage density, while mainstream DRAM today typically ranges between 6F² and 8F².75% similarVerified4-bit量化使模型存储空间缩减为FP16的约1/475% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1057301API
curl https://kongchang.com/api/v1/knowledge/claims/1057301MCP
get_claim(id=1057301)