Unverified50% confidenceBenchmarkExact time
The self-developed W4 kernel's perplexity (probe ppl) is 15.76, better than bitsandbytes NF4's 18.98.
1
Sources
50%
Confidence
Short-term (~14 days)
Relevance
10/12/2026
First Seen
Valid until: 10/26/2026
Sources
Related Entities
Related Claims
UnverifiedIn the same test, the bf16 baseline exceeded the self-developed W4 kernel (25.5 tok/s) and bitsandbytes NF4 (25.7 tok/s) at 31.3 tok/s, reaching 29.8% of the theoretical roofline.73% 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% similarUnverified在4-bit量化下，纯参数理论最小占用为：7B约3.26GiB，13B约6.05GiB，32B约14.90GiB，70B约32.60GiB67% similarUnverified从FP16到Q4_K_M，主流70B模型的困惑度（PPL）提升通常在0.1至0.3区间67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1006841API
curl https://kongchang.com/api/v1/knowledge/claims/1006841MCP
get_claim(id=1006841)