Unverified50% confidenceFactExact time
Meta在Llama 4论文中将MTP作为核心训练目标，DeepSeek-V3也大规模采用了类似机制
1
Sources
50%
Confidence
Long-term
Relevance
7/8/2026
First Seen
Sources
Related Claims
UnverifiedThe core innovations of DeepSeek V4 include continued optimization of the MoE (Mixture of Experts) architecture and deep engineering tuning of the training pipeline.64% similarUnverifiedMTP (Multi-Token Prediction) is a speculative decoding strategy first introduced by DeepSeek in version V3.64% similarUnverifiedDeepSeek introduced MTP technology as early as version V3.64% similarUnverifiedDeepSpeed and Megatron-LM are mainstream large model training frameworks.63% similarUnverifiedDeepSeek V4引入了MHC流形约束连接技术，用于保证深层网络训练稳定性63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/239620API
curl https://kongchang.com/api/v1/knowledge/claims/239620MCP
get_claim(id=239620)