Unverified85% confidenceFactTime unknown
DeepSeek V4采用MUON优化器替代AdamW进行训练
1
Sources
85%
Confidence
Long-term
Relevance
6/2/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedSubsequent versions of DeepSeek V4 will support native multimodality, but multimodality is only one component; the main focus remains on improving AI training and solving continual learning.77% similarUnverifiedThe core innovations of DeepSeek V4 include continued optimization of the MoE (Mixture of Experts) architecture and deep engineering tuning of the training pipeline.76% similarUnverifiedDeepSeek V4 adopted the Muon optimizer improved by the Kimi (Moonshot AI) team, which is a second-order optimization algorithm based on Nesterov momentum.73% similarVerifiedDeepSeek V4 adopts a Mixture of Experts (MoE) architecture.72% similarUnverifiedDeepSeek open-sourced its trained model weights, training data recipe, and technical report.72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/36595API
curl https://kongchang.com/api/v1/knowledge/claims/36595MCP
get_claim(id=36595)