Unverified50% confidenceFactExact time

MoE模型的总参数量与激活参数量之间通常相差10-30倍

1

Sources

50%

Confidence

Long-term

Relevance

8/27/2026

First Seen

Sources

FreeToken：753B大模型塞进单卡的本地部署硬件账该怎么算

Related Entities

Related Claims

VerifiedMoE架构每次推理时通常仅激活总参数的10%-20%75% similar Unverified分析推测该模型是2到4万亿参数规模的MoE模型，每个Token仅激活约1500亿参数71% similar VerifiedMoE架构的单次推理成本相比同等参数规模的稠密模型可降低5-10倍70% similar Unverified若MoE模型每次推理仅激活约1/8的参数，实际计算开销可能仅相当于百亿参数级别的稠密模型69% similar UnverifiedMoE层通常占据模型总参数量的80%以上68% similar

Cite This Claim

Stable URI

https://kongchang.com/claim/809886

API

curl https://kongchang.com/api/v1/knowledge/claims/809886

MCP

get_claim(id=809886)