Unverified85% confidenceFactTime unknown

DeepSeek将后训练阶段的强化学习计算量提升到了预训练的10%左右，而大多数公司这一比例仅为1%-5%

1

Sources

85%

Confidence

Long-term

Relevance

6/1/2026

First Seen

Sources

DeepSeek V3.2深度解读：DSA稀疏注意力、强化学习与Agent三大突破

bilibili网络安全基础官方教程

Related Entities

DeepSeekcompany

Related Claims

UnverifiedDeepSeek将训练成本压缩至业界同类模型的十分之一左右80% similar UnverifiedDeepSeek的训练成本仅为同级别模型的数分之一75% similar UnverifiedDeepSeek 以约 600 万美元训练成本实现竞争力性能，远低于同级别西方模型74% similar UnverifiedDeepSeek在多头潜在注意力（MLA）和FP8混合精度训练方向上的创新将KV Cache内存占用压缩至标准多头注意力的5%-13%73% similar UnverifiedDeepSeek以极低的训练成本引发业界广泛关注72% similar

Cite This Claim

Stable URI

https://kongchang.com/claim/19954

API

curl https://kongchang.com/api/v1/knowledge/claims/19954

MCP

get_claim(id=19954)