Unverified50% confidenceBenchmarkExact time

对照模型Supra-50M有51.8M参数，bf16下103.6MB，8-bit量化后56.2MB，降到int4困惑度从165恶化到193，压到三元直接崩溃

1

Sources

50%

Confidence

Long-term

Relevance

9/15/2026

First Seen

Sources

19.8MB的LLM：44M参数模型如何在CPU上跑出1900 tok/s

Related Entities

Related Claims

Unverified70B模型在INT4精度下可压缩至约35GB74% similar Unverified13B-14B级别模型（4-bit量化）在该配置下速度下降到每秒2-5个token73% similar Unverified2.4万亿参数模型在FP32下约需9.6TB显存，BF16约需4.8TB，INT8约需2.4TB，INT4约需1.2TB72% similar Verified一个70B参数的FP16模型大约需要140GB显存，INT8量化后约需70GB，INT4量化后约需35-40GB71% similar

Cite This Claim

Stable URI

https://kongchang.com/claim/922671

API

curl https://kongchang.com/api/v1/knowledge/claims/922671

MCP

get_claim(id=922671)