Unverified50% confidenceOpinionExact time
NVFP4, with its finer block granularity and floating-point scaling, better preserves small-value information in layers where activation distributions span a wide range, so at the same 4-bit width it often incurs less perplexity loss than INT4.
1
Sources
50%
Confidence
Long-term
Relevance
10/27/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedNVFP4 uses the E2M1 allocation of mantissa and exponent, has an extremely narrow representable range, and therefore must be used with block-wise scaling.79% similarUnverified4-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.73% similarUnverified4-bit quantization means each parameter occupies only 0.5 bytes, reducing VRAM usage by 75% compared to FP16.71% similarVerifiedQ4 means each weight is stored using 4-bit integers, reducing the size by approximately 75% compared to the original 16-bit floating point71% similarUnverifiedINT4 quantization can further compress the KV Cache to 0.5 bytes, but may require calibration data to ensure accuracy.70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1101738API
curl https://kongchang.com/api/v1/knowledge/claims/1101738MCP
get_claim(id=1101738)