Unverified50% confidenceFactExact time
MoE架构在推理时仅激活一小部分专家子网络，2.4T参数模型每次推理可能仅激活约200-400亿参数
1
Sources
50%
Confidence
Long-term
Relevance
7/24/2026
First Seen
Sources
Related Claims
Verified1.6万亿参数级别的巨型模型通常采用混合专家架构（MoE），每次推理只激活一小部分专家子网络80% similarVerifiedMOE架构将模型内部划分为多个专家子网络，每次推理时由门控网络动态选择少数专家参与计算，实际激活参数量远小于总参数量79% similarUnverifiedIn the MoE architecture, a 600-billion-parameter model might only activate 37 billion parameters per inference79% similarUnverifiedMoE架构下模型总参数量远大于单次推理实际激活的参数量，Qwen3.8每次推理激活参数可能仅数百亿级别78% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/604965API
curl https://kongchang.com/api/v1/knowledge/claims/604965MCP
get_claim(id=604965)