deepseek-ai/deepseek-v3

★ 103,409PythonAudience · developerComplexity · 4/5QuietLicenseSetup · hard

Mindmap

mindmap
  root((repo))
    What it does
      671B parameter LLM
      Mixture-of-Experts design
      128K context length
    How it works
      Multi-head Latent Attention
      DeepSeekMoE routing
      Multi-token prediction
      Speculative decoding
    Training details
      14.8T tokens pre-trained
      Supervised fine-tuning
      Reinforcement learning
      R1 reasoning distillation
    Use cases
      Run locally on GPU
      Download from Hugging Face
      Compare against other LLMs
      Build applications
    Tech stack
      Python
      FP8 mixed-precision
      H800 GPUs

mindmap root((repo)) What it does 671B parameter LLM Mixture-of-Experts design 128K context length How it works Multi-head Latent Attention DeepSeekMoE routing Multi-token prediction Speculative decoding Training details 14.8T tokens pre-trained Supervised fine-tuning Reinforcement learning R1 reasoning distillation Use cases Run locally on GPU Download from Hugging Face Compare against other LLMs Build applications Tech stack Python FP8 mixed-precision H800 GPUs

Things people build with this

USE CASE 1

Download and run the model locally on your own GPU hardware to generate text and answer questions.

USE CASE 2

Compare DeepSeek-V3's performance against other large language models like LLaMA and Qwen using the provided evaluation benchmarks.

USE CASE 3

Build applications that use the model's 128K context window to process long documents or conversations.

USE CASE 4

Integrate the model into your own systems using the Hugging Face model hub for inference or fine-tuning.

Tech stack

PythonPyTorchHugging FaceFP8 mixed-precisionH800 GPU

Getting it running

Difficulty · hard Time to first run · 1day+

Requires H800 GPU or equivalent high-end hardware; 671B model needs significant VRAM and specialized inference setup.

Code is licensed under MIT, allowing free use for any purpose including commercial; model weights are under a separate model agreement.

In plain English

DeepSeek-V3 is a large language model released as open source. The README presents it as a Mixture-of-Experts model with 671 billion total parameters, of which 37 billion are activated for each token (a token is a chunk of text the model reads or writes). In a Mixture-of-Experts design, only a slice of the total network is used per word, which keeps the cost of running the model lower than a fully dense model of the same size. According to the README, DeepSeek-V3 reuses architectural ideas from the earlier DeepSeek-V2 called Multi-head Latent Attention and DeepSeekMoE, and adds two new tricks: an auxiliary-loss-free strategy for keeping the experts evenly used, and a multi-token prediction objective during training that the authors say boosts performance and can speed up inference through speculative decoding. The team pre-trained the model on 14.8 trillion tokens, then ran supervised fine-tuning and reinforcement learning stages, and distilled reasoning patterns from a separate DeepSeek-R1 model into V3. Training used an FP8 mixed-precision framework on H800 GPUs and consumed roughly 2.788 million GPU hours total. You would use this repo to download the model weights from Hugging Face and run the model yourself, or to read the technical paper linked inside. The README states a 128K context length and includes evaluation tables comparing DeepSeek-V3 against models such as Qwen2.5 72B and LLaMA3.1 405B, plus instructions for running the model locally further down. The primary language is Python. Code is licensed MIT, while model weights are under a separate model agreement linked in the repo. The full README is longer than what was provided.

Copy-paste prompts

Prompt 1

How do I download and run DeepSeek-V3 locally on my GPU? What are the minimum hardware requirements?

Prompt 2

Explain how the Mixture-of-Experts architecture in DeepSeek-V3 reduces computational cost compared to a fully dense model.

Prompt 3

What is multi-token prediction and how does speculative decoding speed up inference in DeepSeek-V3?

Prompt 4

Show me how to use DeepSeek-V3 from Hugging Face in Python to generate text with a 128K token context.

Prompt 5

How does DeepSeek-V3 compare to LLaMA 3.1 405B and Qwen 2.5 72B on standard benchmarks?

Open on GitHub → Explain another repo

Generated 2026-05-21 · Model: sonnet-4-6 · Verify against the repo before relying on details.