Deploying DeepSeek locally has several key advantages:
- Offline usage with no restrictions.
- Ability to feed custom data and build your own knowledge base.
- Familiarity with the large model deployment process.
- Better understanding of computer configurations.
- Not affected by server limitations.
There are countless online tutorials on deploying DeepSeek, but when attempting it myself, I discovered several important details to keep in mind.
This article will cover the issues I encountered while deploying DeepSeek-R1:70b locally, along with some key insights and experiences.
1. Hardware Requirements
The larger the DeepSeek-R1 model, the more intelligent it becomes, but it also demands higher hardware specifications.
Many online sources have already summarized the hardware requirements. Here, I refer to [1]:
Here is the translated hardware requirements table for DeepSeek-R1:
| Model Version | Model Size | CPU | GPU | RAM | Disk Space |
|---|---|---|---|---|---|
| 1.5B | 1.1GB | Standard quad-core or six-core processor | NVIDIA GTX 1650 or RTX 2060 (mid-range GPU) | 16GB RAM | At least 50GB free space |
| 7B | 4.7GB | 6-core or 8-core processor | NVIDIA RTX 3060 or better | 32GB RAM | At least 100GB free space |
| 8B | 4.9GB | 6-core or 8-core processor | NVIDIA RTX 3060 or better | 32GB RAM | At least 100GB free space |
| 14B | 9GB | 8-core or higher processor (e.g., Intel i9 or AMD Ryzen 9) | NVIDIA RTX 3080 or better | 64GB RAM | At least 200GB free space |
| 32B | 20GB | 8-core or higher processor | NVIDIA RTX 3090, A100, or V100 | 128GB RAM | At least 500GB free space |
| 70B | 43GB | 12-core or higher processor (high-end Intel or AMD recommended) | NVIDIA A100, V100 (may require multiple GPUs) | 128GB RAM | At least 1TB free space |
| 671B | 404GB | High-performance, multi-core CPU (recommended multi-server setup) | NVIDIA A100 or multiple V100 GPUs (cluster support required) | At least 512GB RAM | At least 2TB free space |
It is highly recommended to start with the 1.5B version first. This allows you to successfully run through the entire deployment process before attempting larger versions.
2. Installation Process
The easiest way to deploy DeepSeek is through Ollama:
Here’s how to install DeepSeek on a Linux system:
1. Install Ollama via Command Line
Run the following command in the Terminal:
curl -fsSL https://ollama.com/install.sh | sh
Note: Ollama is approximately 1.4GB, so the download may be very slow.
Important Reminder:
- Do not install Ollama manually.
- It is highly recommended to use the official command-line installation with a global proxy enabled.
Potential Issue:
After manually installing Ollama, I was able to download models, but encountered the following error when running them:
Error: llama runner process has terminated: exit status 2
To check whether Ollama is installed correctly, you can run:
ollama
2. Run Ollama and Download the Model
Open a new Terminal window, then enter the following command:
2. Run Ollama and Download the Model
Start Ollama in the system by running the following command in a new Terminal window:
ollama serve
Now, download the DeepSeek-R1 model (example: 1.5B version) by entering:
ollama pull deepseek-r1:1.5b
3. Run the Model
Once the download is complete, you can run the model using:
ollama run deepseek-r1:1.5b
This will execute DeepSeek directly in the Terminal.
Running DeepSeek in a Web Browser
Running DeepSeek in the Terminal may not be the most convenient option for daily workflow, and it also makes internet access more difficult.
To enhance usability, you can deploy DeepSeek on a webpage and run it in a browser.
Using Page Assist for Web UI
The simplest way to achieve this is by using Page Assist, a web UI designed for Ollama.
- Click “Add to Chrome” to install Page Assist.
- Once installed, go to Chrome Extensions and click on the plugin icon.
- Important: Ensure that Ollama is running before opening the web UI.
- A webpage will pop up, where you can select the DeepSeek model from the dropdown menu.
- Note: Page Assist supports internet access.
Next Steps
In future updates, I will cover how to feed custom data into DeepSeek, integrating it with your own database to make the model more personalized and intelligent. 🚀
