DeepSeek releases DeepSeek-V3
AIMade by AIWritten by AI (Claude) from the source below. Check anything important against the source.How it's made
DeepSeek released DeepSeek-V3 — 671 billion total and 37 billion active parameters per token, and about 2,788,000 H800 GPU hours for the full training run, company-reported. Code under MIT; model weights under the company's own Model License (commercial use allowed)
Why it matters
Drew attention with its claim of extreme training-cost efficiency. The model weights themselves moved to MIT only with DeepSeek-V3-0324 in March 2025
Source — Primary sources first; secondary reports are labelled “Press”
Officialhuggingface.cohttps://huggingface.co/deepseek-ai/DeepSeek-V3Added in Week 43, 2026 · See that week
On the map — ⑨ Open-Weight Lineage line
- Previous stopLlama 3.2
- This recordDeepSeek-V3
- Next stopKanana Nano 2.1B
In the model family tree — DeepSeek family · 12 versions
- Earlier versionDeepSeek-V2
- This recordDeepSeek-V3
- Later versionDeepSeek-R1