Verified85% confidenceFactTime unknown

Mixtral 8x7B模型拥有8个专家，每个Token只激活2个，总参数量达到47B，但单次推理的计算量仅相当于一个13B参数的稠密模型

5

Sources

85%

Confidence

Long-term

Relevance

6/1/2026

First Seen

Sources

NVIDIA Dynamo多轮智能体交互：流式Token与工具调用深度集成

rss

Related Entities

Mixtral 8x7Bmodel MoEconcept

Related Claims

UnverifiedMixtral 8x7B的实际激活参数量仅为总参数量的约1/480% similar UnverifiedMixtral 8×22B 拥有 141B 总参数，但每次推理仅激活约 39B79% similar UnverifiedMistral AI的Mixtral 8x7B采用MoE架构，总参数量约47B，但每次推理实际激活的参数仅约13B75% similar UnverifiedLFM2.5-8B-A1B总参数量为8B，推理时仅激活1.5B参数71% similar

Cite This Claim

Stable URI

https://kongchang.com/claim/21197

API

curl https://kongchang.com/api/v1/knowledge/claims/21197

MCP

get_claim(id=21197)