Unverified50% confidenceBenchmarkExact time
该量化模型相较于BF16全精度版本保留了约77%的准确率
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
8/24/2026
First Seen
Valid until: 11/22/2026
Sources
Related Entities
Related Claims
VerifiedThe larger the number of model parameters, the greater the robustness to quantization, allowing 70B-level models to maintain performance close to FP16 after 4-bit quantization.72% 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70% similarUnverifiedTop-1% accuracy is a metric for evaluating the output consistency between a quantized model and the original full-precision model, focusing on whether the top 1% of tokens with the highest model confidence are consistent with the full-precision version.67% similarUnverifiedUnder the same model (Claude 4.6 Opus), OCR's precision is 33.9% versus 7.23% for the general-purpose agent, representing an approximately 4.7× improvement in precision.65% similarUnverifiedModels like GPT-4 and Claude 3.5 surpassed 90% accuracy on MMLU, significantly diminishing its discriminative power64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/795615API
curl https://kongchang.com/api/v1/knowledge/claims/795615MCP
get_claim(id=795615)