Skip to content

DeepSeek-V3

A 671B-parameter MoE language model

Visit official site

What DeepSeek-V3 is

The model is described as a Mixture-of-Experts language model with 671B total parameters and 37B activated at a time.

Model notes

  • Pre-trained on 14.8 trillion high-quality tokens
  • Auxiliary-loss-free load balancing and multi-token prediction
  • FP8 mixed-precision training framework

Application scenarios

  • Analyse financial news and reports for key information
  • Understand medical literature for drug and case research
  • Help students and researchers find academic material

Getting started

  • Clone the GitHub repository and install requirements.txt
  • Download model weights from HuggingFace
  • Convert FP8 weights to BF16 with the provided script if needed

Pros & cons

✓ What we like

  • FP8 mixed-precision training framework
  • Model weights offered on HuggingFace

! What to watch out for

  • Local setup requires the repository and its dependencies
  • FP8 weights may need conversion to BF16

FAQ

How large is the model?

671B total parameters and 37B activated at a time.

Where are the weights?

A 685B HuggingFace download includes 671B main weights and a 14B module.

Last reviewed: 2026-11-15

More AI coding tools tools

View all →

How we review