DeepSeek-R1

DeepSeek-R1

Generates text with a 684B reasoning model trained mainly through reinforcement learning, aimed at math, code and logic. Suits researchers and teams who host large open models.

Made by deepseek-ai on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: DeepSeek-R1 is a 684B parameter reasoning model from DeepSeek. It builds on DeepSeek-R1-Zero, a model trained with large-scale reinforcement learning and no supervised fine-tuning step first. R1 adds cold-start data before the reinforcement learning, which the card says fixes repetition, poor readability and mixed languages. Its stated results match OpenAI-o1 on math, code and reasoning tasks.

Who it suits: Research groups and engineering teams that can host a huge open model and want strong step-by-step reasoning. People with a single desktop GPU will want one of the smaller distilled models instead.

Industry | Who uses it | How they use it | Example
Research | AI labs | Study reasoning trained through reinforcement learning | Compare R1 with R1-Zero on reasoning tasks
Software | Platform teams | Serve a large open reasoning model for coding and math | Host it behind an internal chat tool
Education | Tutors and students | Work through math and logic problems step by step | Ask it to solve and explain a proof

If you cannot host about 684 GB of weights, you can skip it.

Where to use it:

• Python with the transformers library
• Large GPU servers, since the weights are about 684 GB in F8_E4M3
• Windows, macOS or Linux, for downloading the files

Use case | Who | What they get back
Math problems | Students and researchers | A worked solution with reasoning
Code tasks | Developers | Code and an explanation of the approach
Reasoning tasks | Analysts | A chain of reasoning ending in an answer

How to set it up:

• Python with pip
• Huge storage and GPU memory, about 684 GB for the weights and about 342 GB at 4-bit by Hugging Face's estimate
• A read of the Usage Recommendation section of the model card before running it locally

1. Install the Hugging Face hub package.
2. Download the model files.
3. Review the card's usage recommendations, then run it with your serving stack.

pip install -U huggingface_hub
hf download deepseek-ai/DeepSeek-R1

How to use it: Send the model a math, code or reasoning problem through your serving stack. The card asks you to read its Usage Recommendation section first, because the R1 series has settings meant for running it well.

Example: You ask a hard math question. The model reasons through the steps at length and then states its answer.

For newcomers and experts: If you are new, a 684B model is far too large for a normal computer, so try one of the six smaller distilled models from the same release. Use R1 itself when you run a GPU cluster.

The training pipeline has two reinforcement learning stages that look for better reasoning patterns and match human preferences, plus two supervised fine-tuning stages that seed reasoning and non-reasoning skills. The card says the six distilled models are built on Llama and Qwen, and its 32B Qwen distill beats OpenAI-o1-mini on many benchmarks.

Base model
Reinforcement learning builds reasoning
Cold-start data and fine-tuning clean it up
A second reinforcement learning stage aligns it
DeepSeek-R1

Good to know:

• The weights are about 684 GB, so it needs serious hardware
• DeepSeek-R1-Zero showed that reasoning can come from reinforcement learning alone, with some repetition and language mixing
• License: MIT
• Cost to run: none named, but the hardware is large
• Data it reaches: your prompts and the files downloaded from Hugging Face
• Not for: ordinary laptops, where the distilled models fit better

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 684B
• Weights: about 684 GB in F8_E4M3; about 342 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 1,599,170
• Likes: 14,451
• License: mit
• Quantized versions: 70
• Fine-tunes: 328
• Tags: transformers, safetensors, deepseek_v3, text-generation, conversational, fp8

Links

• Model: https://huggingface.co/deepseek-ai/DeepSeek-R1
• Files: https://huggingface.co/deepseek-ai/DeepSeek-R1/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download deepseek-ai/DeepSeek-R1

Then follow the usage in its model card. Nothing in this kit runs by itself.