Unverified50% confidenceFactExact time
DeepSeek-V2将每个Transformer层的FFN拆分为多个专家模块，通过门控网络动态路由，每个token只激活约10%的参数
1
Sources
50%
Confidence
Long-term
Relevance
8/24/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedDeepSeek采用MoE（混合专家模型）架构，与传统密集型Transformer模型每次推理激活全部参数不同，MoE每次推理由路由器动态选择少数专家参与计算67% similarVerifiedDeepSeek uses a Mixture of Experts (MoE) architecture, activating only a subset of expert sub-networks rather than all parameters during each inference.65% similarVerifiedDeepSeek-V3 has 671B parameters but activates only 37B, adopting an architecture of 256 routed experts plus 1 shared expert, with each token processed by 8 selected experts.62% similarUnverifiedThe combination of DeepSeek + Dify + Ollama forms a complete privatized technology stack covering everything from model inference to application orchestration.62% similarUnverifiedThe ranking layer commonly uses deep learning models such as Wide&Deep, DeepFM, or Transformer architectures for CTR prediction and interaction probability calculation.62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/795622API
curl https://kongchang.com/api/v1/knowledge/claims/795622MCP
get_claim(id=795622)