Unverified50% confidenceFactExact time

FP4 相对 FP16 可节省75%存储，采用每组16或32个权重共享一个 FP8 缩放因子的分组量化策略

1

Sources

50%

Confidence

Long-term

Relevance

7/8/2026

First Seen

Sources

Unsloth重磅更新：支持NVFP4量化导出与DeepSeek-V4训练

rss7/7/2026

Related Claims

Partially Verified标准的FP32模型每个参数占用4字节，而INT4量化将其压缩至0.5字节，理论上可实现8倍的内存节省77% similar UnverifiedBlackwell将FP4引入推理，配合微缩放技术可将权重存储压缩至FP16的1/474% similar UnverifiedFP8相比BF16可提供约2倍的内存带宽节省和更高的矩阵计算吞吐72% similar UnverifiedINT4 相较 FP16 实现了 4 倍存储压缩72% similar

Cite This Claim

Stable URI

https://kongchang.com/claim/295842

API

curl https://kongchang.com/api/v1/knowledge/claims/295842

MCP

get_claim(id=295842)