DeepSeek-V3 Technical Report

DeepSeek-V3 Technical Report — a 671B/37B MoE trained in FP8

Why it matters

Established low-precision pretraining at frontier scale

Source — Primary sources first; secondary reports are labelled “Press”

Paperarxiv.orghttps://arxiv.org/abs/2412.19437

Added in Week 35, 2026 · See that week

Same week

Geoffrey Hinton raises the odds of AI wiping out humanity in 30 years to 10–20%PressNational Assembly passes Korea's AI Basic Act in plenary sessionPressxAI raises a $6 billion Series COfficial

Nearby on the same axis

Byte Latent TransformerPaperKeller Jordan releases the Muon optimizerOfficialNie et al. present LLaDAPaperNSA — Native Sparse AttentionPaper