DeepSeek · LLMs

DeepSeek-V3

Name: DeepSeek-V3
Author: DeepSeek

DeepSeek's 671B-parameter mixture-of-experts model with 37B active per token, notable for frontier-level performance at a reported low training cost.

Official page

Run it locally

With Ollama installed, download and start the default version with:

ollama run deepseek-v3:671b

Variant	Download
deepseek-v3:671b	404GB

Memory: About 404GB for the default deepseek-v3:671b version (4-bit)
Formats: Safetensors (Hugging Face), GGUF (Ollama, llama.cpp, LM Studio)

The download size is roughly the memory (RAM or GPU VRAM) needed to load a variant; allow extra for long contexts. See our hardware guide.

Overview

DeepSeek-V3, released in December 2024, uses a mixture-of-experts architecture with multi-head latent attention. Its technical report described training on 14.8 trillion tokens with a comparatively small compute budget, which drew wide attention.

Highlights

671B total, 37B active parameters
128K-token context window
Multi-token prediction and FP8 training
Basis for the DeepSeek-R1 reasoning model

Typical uses

Self-hosted assistants, coding and analysis; widely available through cloud providers.

Things to consider

Very large: local use needs around 404 GB of memory.
Data handling of the official app and API is subject to Chinese jurisdiction; self-hosting avoids this.
Download sizes refer to the default 4-bit quantized versions in the Ollama library; full-precision weights are larger.

Specifications are taken from the model card on Hugging Face, the Ollama library and the provider’s announcement. Benchmark results depend on evaluation settings; see how AI benchmarks work.

#LLMs #Open Weights #Local AI #DeepSeek