Unverified50% confidenceFactExact time
DeepSeek的细粒度专家拆分创新是把专家数从N增加到2N甚至更多，使路由选择组合空间从C(8,2)=28种扩展到C(64,6)=7461万余种
1
Sources
50%
Confidence
Long-term
Relevance
9/9/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedDeepSeek的关键创新在于细粒度专家分割策略，将传统MoE中较大的专家拆分为更多更小的单元77% similarUnverifiedDeepSeek正在研发一个规模更大的前沿模型，意在角逐万亿参数量级69% similarUnverified现代深度网络往往拥有数十亿乃至上万亿参数，远超训练样本数量，按经典统计学习理论应严重过拟合，实际上却展现出惊人的泛化能力65% similarVerifiedDeepSeek-V3拥有671B参数但仅激活37B，采用256个路由专家加1个共享专家的架构，每个token选择8个专家处理65% similarUnverifiedDeepSeek与美国相比在算力上差一个数量级还不止64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/885821API
curl https://kongchang.com/api/v1/knowledge/claims/885821MCP
get_claim(id=885821)