Unverified50% confidenceFactExact time
In the decode phase, DeepSeek spreads MoE experts across more than 64 GPUs, with each GPU holding only a small number of experts.
1
Sources
50%
Confidence
Long-term
Relevance
10/21/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedFor MoE, DeepSeek keeps a spare slot on each GPU for a copy of a hot expert, and the system moves these copies every few thousand steps based on traffic conditions.74% similarUnverifiedDeepSeek separates the pre-fill and decode phases onto different machines: pre-fill runs on regular eight-GPU machines, while decode runs on a supernode composed of hundreds of GPUs.73% similarUnverifiedThe MoE structure of DeepSeek-V4.1-Flash has 40 layers, each layer has 384 routed experts, and each token activates only 6 experts per layer.69% similarVerifiedDeepSeek released DeepSeek-V4.1-Flash, a method that significantly compresses the memory required for KV cache.66% similarUnverifiedDeepSeek目前只有约两万张H系列等效GPU算力65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1082058API
curl https://kongchang.com/api/v1/knowledge/claims/1082058MCP
get_claim(id=1082058)