Unverified50% confidenceFactExact time
Google 2017年提出的Sparsely-Gated MoE首次将稀疏门控引入Transformer规模的序列模型，可在不显著增加计算量前提下将模型容量扩大1000倍以上
1
Sources
50%
Confidence
Long-term
Relevance
7/24/2026
First Seen
Sources
Related Claims
UnverifiedGoogle 的 Switch Transformer（2021）和 Mistral 的 Mixtral 8x7B（2023）都采用了稀疏 MoE 架构，即每次前向传播只激活少数几个专家65% similarUnverifiedSwitch Transformer and the Mixtral series have widely adopted the MoE architecture.60% similarUnverifiedBlackwell's second-generation Transformer Engine adds FP4 precision support, which in theory can improve Transformer model training throughput by 2–4x compared to the H100.60% similarUnverifiedGShard于2020年首次在Transformer中实现了600B参数的MoE训练59% similarUnverifiedDiT（Diffusion Transformer）架构的参数量已达数十亿级别，单次推理的中间激活值在高分辨率下可占满80GB A100的全部显存59% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/604998API
curl https://kongchang.com/api/v1/knowledge/claims/604998MCP
get_claim(id=604998)