Deploying DeepSeek locally can be a rewarding experience, allowing you to harness the power of this advanced AI model without relying on external servers. This guide will walk you through the process of setting up DeepSeek on your local machine, ensuring you can make the most of its capabilities.
Understanding the DeepSeek Model
Before diving into the setup, it’s important to understand the differences between the official DeepSeek R1 model and the locally deployable versions. The official R1 model is the “complete version,” utilizing MLA and MoE architectures with 671 billion parameters, activating 37 billion during inference. It has been trained using the GRPO reinforcement learning algorithm.
DeepSeek-R1-1.5B – the best version of DeepSeek I recommend
What I recommend is the DeepSeek-R1-1.5B version. Why? Because it’s a lightweight model with only 1.5 billion parameters. Sounds pretty “mini,” right? But don’t underestimate it—it’s a “small but powerful” model. It only requires 3GB of GPU memory to run, which means even if your computer isn’t high-end, you can easily handle it. Plus, it performs exceptionally well in mathematical reasoning, even surpassing GPT-4 and Claude 3.5 in some benchmark tests.
steps for locally deploying DeepSeek.
Step 1: Choose Your DeepSeek Model
To begin, decide which version of the DeepSeek model you want to deploy locally. I recommend is the DeepSeek-R1-1.5B version. Of course, if your computer has higher specs, you can try other versions.
Step 2: Download and Install Ollama – a large language model tool that supports DeepSeek
Ollama is a tool for running and managing large language models (LLMs) locally. It simplifies the process of setting up and running these models, which can be complex and resource-intensive.
Ollama allows users to run DeepSeek models locally, avoiding the latency and costs associated with using cloud services.
Visit the official website: https://ollama.com/ to download.
The installation file, OllamaSetup.exe, is approximately 745MB.
Double-click the OllamaSetup.exe to begin the installation.
Step 3: Verify if Ollama is installed successfully.
In the command line, type the command ollama -v. If the version number appears as shown, it means the installation was successful.
Step 4: Pull the DeepSeek model through Ollama.
To pull the DeepSeek model through Ollama, you can use the following command in the command line:
ollama run deepseek-r1:1.5b
This will initiate the download of the DeepSeek model via Ollama. Make sure that you have a stable internet connection during the process.
After 5 to 6 minutes, seeing the word “success” indicates that DeepSeek R1 has been successfully installed, and you can then start conversing with DeepSeek.
DeepSeek 1.5B – the entire model size is 1.1 GB.
- 1.5B: Suitable for lightweight tasks, such as simple interactions on edge devices (like smartwatches, IoT devices), small-scale intelligent Q&A systems, etc. Currently, the smallest open-source version.
- 671B: Primarily used for large-scale cloud inference, suitable for research analysis, data mining, and other complex tasks that require processing massive amounts of data. Currently, the most powerful open-source version.
- More versions can be found here: https://ollama.com/library/deepseek-r1.
Step 5: Install WebUI – Interact with DeepSeek through a graphical interface
The installed DeepSeek can interact using the command prompt. However, using the command prompt to converse with DeepSeek is not very user-friendly. For a better experience, we can install a WebUI, and here we use the browser plugin: Page Assist. After starting the Ollama service, press the shortcut key [ctrl + shift + L] to open the WebUI page.
Hardware Requirements for Different Versions of the DeepSeek Model
| Model Version | Parameters | Memory Requirement (FP16) | Recommended GPU (Single Card) | Multi-Card Support | Quantization Support | Suitable Scenarios |
|---|---|---|---|---|---|---|
| DeepSeek-R1-1.5B | 1.5 Billion | 3GB | GTX 1650 (4GB VRAM) | Not Required | Supported | Low-resource device deployment (Raspberry Pi, old laptops), real-time text generation, embedded systems |
| DeepSeek-R1-7B | 7 Billion | 14GB | RTX 3070/4060 (8GB VRAM) | Optional | Supported | Medium-complexity tasks (text summarization, translation), lightweight multi-turn dialogue systems |
| DeepSeek-R1-8B | 8 Billion | 16GB | RTX 4070 (12GB VRAM) | Optional | Supported | Tasks requiring higher precision (code generation, logical reasoning) |
| DeepSeek-R1-14B | 14 Billion | 32GB | RTX 4090/A5000 (16GB VRAM) | Recommended | Supported | Enterprise-level complex tasks (contract analysis, report generation), long text understanding and generation |
| DeepSeek-R1-32B | 32 Billion | 64GB | A100 40GB (24GB VRAM) | Recommended | Supported | High-precision specialized tasks (medical/legal consulting), multi-modal task preprocessing |
| DeepSeek-R1-70B | 70 Billion | 140GB | 2x A100 80GB/4x RTX 4090 (multi-card parallel) | Required | Supported | Research institutions/large enterprises (financial forecasting, large-scale data analysis), high-complexity generation tasks |
| DeepSeek-671B | 671 Billion | 512GB+ (Extremely high single-card memory demand, usually requires multi-node distributed training) | 8x A100/H100 (server cluster) | Required | Supported | National-scale/mega-scale AI research (climate modeling, genome analysis), general AI (AGI) exploration |
Practical Application Scenarios: What Can DeepSeek-R1-1.5B Do?
Despite its small size, DeepSeek-R1-1.5B is not “weak” at all. It is very suitable for lightweight tasks such as:
- Intelligent Customer Service: In small businesses or personal projects, it can quickly answer common customer questions, improving service efficiency.
- Language Learning: You can use it to practice language expression, such as inputting a Chinese sentence and having it generate an English translation.
- Creative Writing: If you are a writer or copywriter, it can help you quickly generate creative snippets or draft copy.
Conclusion
Do you think deploying the DeepSeek-R1-1.5B model locally is super simple? By following the steps above, you can have a “smart assistant” on your computer.
Moreover, this model not only runs quickly but can also play a significant role in many scenarios.
Deploying DeepSeek locally offers a unique opportunity to explore its advanced features without the limitations of server constraints. By following this guide, you can set up and run DeepSeek on your own hardware, unlocking its potential for a wide range of applications.
