Skip to content
Get packs

Guide · Open source

How to use DeepSeek-V3: download, run locally and what you need

Updated · written with AI from the project's own documentation and checked against our catalogue data

DeepSeek-V3 is a large open-weight language model, a text-generating AI system with 671B total parameters, of which 37B are activated for each token. This guide explains what the GitHub repository actually contains, where the model weights are published, and what it takes to run the model yourself. It is written for readers who want to decide whether self-hosting this model is realistic before they spend time on it.

DeepSeek-V3deepseek-ai/DeepSeek-V3 · ★ 105k · MIT

Quick start

  1. 1Decide if you need the model at all. If you only want to chat, use the hosted chat site linked in the README. Self-hosting only makes sense if you have the hardware and a reason to run the weights yourself.
  2. 2Clone the repository. Run the git clone command from the README to get the model card, README_WEIGHTS.md and the license files.
  3. 3Read the license files. Check LICENSE-CODE for the MIT terms and LICENSE-MODEL for the Model Agreement that covers the weights.
  4. 4Check your storage and hardware. The Hugging Face download totals 685B per the README, so confirm you have the storage and GPU capacity before downloading.
  5. 5Download the weights from Hugging Face. Use the deepseek-ai/DeepSeek-V3 link in the model downloads table. Choose the base or instruct variant that fits your goal.
  6. 6Follow Section 6 for local running. Open the How to Run Locally section of the README and follow the community or vendor route that matches your hardware.

What is in the DeepSeek-V3 repository?

The repository is a model card with benchmark results rather than a finished application. The README describes DeepSeek-V3 as a Mixture-of-Experts (MoE) language model, and it links to the model weights on Hugging Face, a chat website and an API platform.

If you only want to chat with the model, the README points to a hosted chat site, so you do not need this repository for that.

  • Model weights: published on Hugging Face under deepseek-ai/DeepSeek-V3 and deepseek-ai/DeepSeek-V3-Base.
  • Chat website and API platform: linked from the README.
  • Benchmarks: a table comparing DeepSeek-V3 with DeepSeek-V2, Qwen2.5 72B and LLaMA3.1 405B.

How do you get the code and model files?

The README's install step is to clone the repository. That gives you the model card, the README_WEIGHTS.md file and the license files. The weights themselves are downloaded separately from Hugging Face.

The README notes that the total size of DeepSeek-V3 on Hugging Face is 685B, which includes 671B of main model weights and 14B of Multi-Token Prediction (MTP) Module weights. Plan storage accordingly before you start a download.

  • Install method: clone the repository from GitHub.
  • Weights: fetch from Hugging Face after cloning.
  • Total download size per the README: 685B parameters' worth of weights.
git clone https://github.com/deepseek-ai/DeepSeek-V3

What hardware do you need to run it locally?

The README does not list minimum hardware or prerequisites. It does say the model is 671B parameters, and the Hugging Face download totals 685B. That scale is far beyond a typical desktop or single consumer GPU.

The README says the project partnered with open-source communities and hardware vendors to offer multiple ways to run the model locally, and that Section 6 gives step-by-step guidance. Read that section for the supported routes rather than assuming a single-machine setup.

  • Expect large GPU infrastructure; the README does not give a specific GPU count or memory figure.
  • Check Section 6 of the README for the community and vendor options it links to.
  • MTP support is described in the README as under active development in the community.

What does the model do well, according to the README?

The README reports that DeepSeek-V3 was pre-trained on 14.8 trillion tokens and then went through Supervised Fine-Tuning and Reinforcement Learning. It states that the model outperforms other open-source models and achieves performance comparable to leading closed-source models on its evaluations.

The architecture uses Multi-head Latent Attention (MLA) and DeepSeekMoE, and the README says the model can also be used for speculative decoding to speed up inference. These are the README's own claims; verify them against the benchmark tables and papers before relying on them.

  • Architecture: Mixture-of-Experts with 37B activated parameters per token.
  • Context length: 128K for both listed models.
  • Multi-Token Prediction can also be used for speculative decoding, per the README.

Which licence applies?

The code and the model weights have different licences. The code is under MIT, but the model weights fall under a separate Model Agreement. Read LICENSE-MODEL before any commercial or redistribution use.

If you need a fully permissive licence that covers the weights, this model may not fit your use case, and a different open-source model may be a better choice.

  • Code licence: MIT (LICENSE-CODE).
  • Model licence: Model Agreement (LICENSE-MODEL).

What people use it for

  • Researchers comparing open-weight large language models against closed commercial models on the README's listed benchmarks.
  • Developers who want to study the Mixture-of-Experts architecture and Multi-Token Prediction design.
  • Teams evaluating open-source options for self-hosting when they have large GPU infrastructure.
  • Anyone reading the benchmark tables to compare DeepSeek-V3 with Qwen2.5 72B and LLaMA3.1 405B.

Frequently asked questions

Is DeepSeek-V3 free to use?

The code is under the MIT licence, but the model weights fall under a separate Model Agreement. Read LICENSE-MODEL before using the weights commercially or redistributing them. The README does not list any prices for the hosted chat or API options.

What are the requirements to run DeepSeek-V3?

The README does not list minimum hardware or prerequisites. The model has 671B total parameters and the Hugging Face download totals 685B, so it needs large GPU infrastructure. Section 6 of the README links to community and vendor routes for running it locally.

How do I install DeepSeek-V3?

The README gives one install command: git clone https://github.com/deepseek-ai/DeepSeek-V3. The model weights are then downloaded from Hugging Face. The repository itself is a model card, not an installable app.

Can I just chat with DeepSeek-V3 without installing it?

Yes. The README links to a chat website and an API platform. For ordinary chatting, use the hosted site instead of running the weights yourself.

How does DeepSeek-V3 compare with Qwen2.5 or Llama 3.1?

The README includes a benchmark table comparing DeepSeek-V3 with Qwen2.5 72B and LLaMA3.1 405B, where DeepSeek-V3 scores highest on many but not all metrics. Those are the authors' results. Test on your own tasks before choosing.

Why does the model licence matter?

Because the weights are under the Model Agreement, not MIT. If you need a fully permissive licence for any use, a model with an OSI-style licence for its weights may suit you better.

New agent skills, repos and AI tools — every week

The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.

Our verdict

DeepSeek-V3 is worth a look for researchers and developers who have large GPU infrastructure and want to study or self-host a very large open-weight model. Skip it if you only want to chat, lack serious hardware, or need weights under a fully permissive licence, where a smaller model such as those in the comparison list may be a better fit.

Install options & downloads