返回文章列表
Models15 min
DeepSeek V3.2 Architecture Deep Dive: MLA, MoE and FP8 Training
Shota TanakaAI Research Analyst2026-04-2415 min
DeepSeekMoEFP8 TrainingMLALLM Architecture
本文以日语发表。中文摘要如下：
DeepSeek V3.2 Architecture Deep Dive: MLA, MoE and FP8 Training—Deep dive into DeepSeek V3.2: 671B-param MoE with 37B active, MLA attention, auxiliary-loss-free load balancing, and FP8 native training — explained from the paper and public implementation.