Unverified50% confidenceFactExact time
The MoE model backbone outputs MoeModelOutputWithPast, and the model head outputs the MoE causal language model output class.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/6/2026
First Seen
Valid until: 1/4/2027
Sources
Related Entities
Related Claims
VerifiedMoE (Mixture of Experts) uses dynamic routing to have different parameter sub-networks handle different types of inputs, greatly increasing model capacity without linearly increasing inference costs.72% similarUnverifiedIn MoE models, each token is routed to only a small number of experts, producing two communication primitives: Dispatch and Combine69% similarUnverifiedKimi系列模型底层架构采用了混合专家模型（MoE）设计67% similarUnverifiedThe core idea of the Mixture of Experts (MoE) model is to activate only a subset of all parameters for a given input, with a gating network dynamically determining the routing path.67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/981733API
curl https://kongchang.com/api/v1/knowledge/claims/981733MCP
get_claim(id=981733)