Unverified50% confidenceFactExact time
The researchers' reference runs use 4-bit NF4 quantization, so it is impossible to determine how much of the reported gains would remain under matched reading and precision conditions.
1
Sources
50%
Confidence
Long-term
Relevance
10/20/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedRe-quantizing INT4 weights to NF4 subjects each weight to two rounds of rounding, and official measurements show this increases the perplexity of Qwen3-8B by 1.6%70% similarUnverifiedNVFP4量化方案相比BF16基线实现了1.5倍的推理加速，同时在top-1准确率上保留了92%到97%的水平68% similarUnverifiedThe self-developed W4 kernel's perplexity (probe ppl) is 15.76, better than bitsandbytes NF4's 18.98.68% similarUnverifiedLoading a model with N parameters at FP16 precision requires approximately 2N bytes of video memory.68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1074072API
curl https://kongchang.com/api/v1/knowledge/claims/1074072MCP
get_claim(id=1074072)