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

Unsloth v0.1.45发布：Gemma 4 MTP加速、全平台修复与Studio全面升级

rss6/11/2026

Related Claims

UnverifiedDeepSeek V4的核心创新包括对MoE（混合专家）架构的持续优化和对训练流水线的深度工程调优64% similar UnverifiedMTP（多令牌预测）是一种投机解码策略，由DeepSeek在V3版本中首次引入64% similar UnverifiedDeepSeek早在V3版本就引入了MTP技术64% similar UnverifiedDeepSpeed、Megatron-LM 是主流的大模型训练框架63% similar UnverifiedDeepSeek 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/239620

API

curl https://kongchang.com/api/v1/knowledge/claims/239620

MCP

get_claim(id=239620)