Unverified50% confidenceFactExact time
DeepSeek发展出MLA（多头潜在注意力）等节省显存的创新技术
1
Sources
50%
Confidence
Long-term
Relevance
7/25/2026
First Seen
Sources
梁文锋4小时内部对话：DeepSeek的克制哲学与AGI路线
Related Claims
VerifiedDeepSeek由深度求索公司开发79% similarVerifiedDeepSeek adopts a Mixture of Experts (MoE) architecture and a Multi-head Latent Attention (MLA) mechanism.79% similarVerifiedDeepSeek introduced innovations such as auxiliary-loss-free load balancing in its MoE routing strategy, improving expert utilization.78% similarUnverifiedDeepSeek以技术报告作为主要发布媒介推出模型77% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/613494API
curl https://kongchang.com/api/v1/knowledge/claims/613494MCP
get_claim(id=613494)