Unverified60% confidenceFactExact time
The MoE architecture was founded on a Google paper from 2017; it dynamically assigns tokens to a small number of expert sub-networks via a routing network, and Qwen 3.5 122B has 122B total parameters but activates only approximately 22B during each inference.
2
Sources
60%
Confidence
Long-term
Relevance
7/12/2026
First Seen
Sources
AMD Ryzen AI Halo Hands-On: Can x86 Architecture Challenge NVIDIA's DGX Spark?
bilibili量子菠萝_7/7/2026
Related Claims
UnverifiedMoE架构由多个专家子网络和一个路由器组成，路由器根据输入token动态选择少数几个专家参与计算，DeepSeek V2/V3、Mixtral 8x7B、Google Gemini 1.5均采用这一架构72% similarUnverifiedThe MoE architecture divides the model into multiple expert sub-networks, activating only a small portion during each inference, reducing computational costs while maintaining knowledge capacity; Google's Gemini and Mistral's Mixtral series both adopt this architecture.69% similarUnverifiedMoE架构由 Google Brain 于 2017 年提出，GPT-4、Mixtral 8x7B 等知名模型均采用该架构68% similarVerifiedThe Mixture of Experts (MoE) architecture was scaled up within the Transformer framework by Google Brain in 2017; its core idea is to group model parameters into multiple expert sub-networks, with only a small subset activated during each inference.67% similarUnverifiedThe MoE (Mixture of Experts) architecture was first proposed by Hinton's team in 2017, dynamically selecting Top-K expert sub-networks to participate in computation through a gated routing network.67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/489932API
curl https://kongchang.com/api/v1/knowledge/claims/489932MCP
get_claim(id=489932)