Unverified50% confidenceFactExact time
三值化权重每个权重仅携带约1.58 bits信息量（log₂3≈1.585），相比FP32的32 bits信息密度下降约20倍
1
Sources
50%
Confidence
Long-term
Relevance
7/17/2026
First Seen
Sources
Related Claims
UnverifiedA 140 GB FP16 70B model can be compressed to approximately 35–40 GB after 4-bit quantization, reducing communication overhead by about 75%; 2-bit quantization can compress it to approximately 18 GB but introduces accuracy loss.74% similarUnverifiedTaking the Q4_0 format as an example, every 32 FP32 weights are packed into one quantization block containing one scale factor and 32 4-bit integers, averaging approximately 4.5 bits per parameter, reducing memory usage to approximately 1/8 of the original.72% similarUnverifiedEven compressing the weights to 214GB using 4-bit quantization still far exceeds the capacity of a single 48GB GPU.71% similarUnverifiedThe Bonsai ternary version is approximately 6GB at 1.71 bits and retains approximately 97% of the base model's quality; the 1-bit version is 3.9GB and retains approximately 90% of quality69% similarUnverifiedCalculated from a storage perspective: 27B parameters × 2 bytes (FP16) ≈ 54 GB, INT8 ≈ 27 GB, INT4 ≈ 13.5 GB, 2-bit ≈ 6.75 GB, 1.5-bit approaches 5 GB; fitting within 4 GB means the average bit width needs to be below 1.5-bit.69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/538892API
curl https://kongchang.com/api/v1/knowledge/claims/538892MCP
get_claim(id=538892)