DeepSeek R1, a new open-source AI language model developed by the Chinese AI company DeepSeek, is popular right now.
Some users claim that it is comparable to or even stronger than OpenAI’s o1 in terms of inference capabilities.
Currently, DeepSeek is free, which is good news for users, but it also raises some questions.
How will they handle the cost of the server as the number of users increases?The hardware running the model shouldn’t be cheap, right?
One reasonable guess is data.
Data is the lifeblood of AI models.They may collect user data in some way, and this data may be useful for their trading models or other forms of monetization.
If you’re worried about data privacy but still want to use R1 without sharing data, the best option is to run the model locally.
What is DeepSeek R1?
A few days ago, DeepSeek R1 was introduced as a fully open source model. This means that anyone can access its underlying code, make modifications, and even fine-tune it to their own needs.
DeepSeek R1 (often abbreviated as R1) is derived from a large base model called DeepSeek-V3.
The research team has optimized this model by combining supervised fine-tuning (SFT) with high-quality manually annotated data and reinforcement learning (RL).
This has resulted in a chatbot that can handle complex prompts, show the reasoning behind complex questions (sometimes more clearly than other models), and even present code in the chat interface for quick testing.
It’s really impressive, especially for an open-source model.
How to run DeepSeek R1 locally
To run DeepSeek R1 locally, we will use a tool called Ollama.
Ollama is a free and open source tool that allows users to run large language models (LLMs) locally on their computers.
It is available for macOS, Linux, and Windows.Visit the official Ollama website, click the “Download” button, and install it on your system.
To confirm a successful installation,
open a terminal and run the following command:ollama -v
You should see the Ollama version number instead of an error message.
Under the “Models” tab, search for the keyword “deepseek” and you should see “deepseek-r1” at the top of the list.
Click on it, and in the “Models” section, you will see that there are multiple model sizes with parameters ranging from 5 billion to 671 billion.As a general rule, larger models require more powerful GPUs to run.
A small model like the 8 billion parameter version can run on a GPU with 8 GB of graphics memory, while larger models require significantly more resources (see the section below on graphics memory and GPU requirements).
To download and run the 8 billion parameter model, use the following command:ollama run deepseek-r1:8b
The model will start downloading (approx. 4.9 GB).Before proceeding, please ensure that you have sufficient disk space.
Once the download is complete, the model will be running on your local computer, and you can start a conversation with it immediately.
Let’s test this with a sample prompt:
Prompt: What is DeepSeek R-1?
Response: DeepSeek R-1 is an AI assistant developed by the Chinese company DeepSeek.
It is designed to provide answers and help on a variety of topics, including but not limited to mathematics, programming, natural language processing, etc. If you have any questions or need help, feel free to ask me!
Great.
It is fast and works even when I disconnect my laptop from the Wi-Fi network. Please note that it cannot access the internet even if you are connected to it.
Tip: What is the latest share price of AAPL?
Response: As an AI assistant, I cannot access real-time data and therefore cannot provide the latest share price of Apple Inc. (AAPL).
For the most accurate and up-to-date information, I suggest you check financial news platforms or your brokerage services.Other things Ollama can do:
Run LLMs locally, including LLaMA2, Phi 4, Mistral and Gemma 2Allow users to create and share their own LLMsPackage model weights, configurations, and data into a single packageOptimize settings and configuration details, including GPU usage.
GPU and memory requirements
The memory requirements for DeepSeek-R1 depend on factors such as the size of the model, the number of parameters, and the quantization technique.
Here is a detailed overview of the memory requirements for DeepSeek-R1 and its reduced model, as well as the recommended GPUs:
Key note on memory usage:
Distributed GPU setup for large models: DeepSeek-R1-Zero and DeepSeek-R1 require a lot of graphics memory and therefore a distributed GPU configuration (e.g. NVIDIA A100 or H100 in a multi-GPU setup) for optimal performance.
Lite models are optimized to run on a single GPU with lower graphics memory requirements, starting at 0.7 GB.
Additional memory usage: Activation, buffers, and batch tasks may consume additional memory.
Why run locally?
DeepSeek’s web chatbot and mobile app are free and very convenient. You don’t need to do any setup, and features like DeepThink and web search are built-in.
However, running it locally may be a better option for the following reasons
Privacy
When you use the web or app version, your queries and any attached files are sent to DeepSeek’s servers for processing.What happens to this data?We don’t know.Running the model locally ensures that your data stays on your computer, giving you complete control over your privacy.
Offline access
Running the model locally means you don’t need an internet connection.
If you’re traveling, encountering unstable Wi-Fi, or simply prefer working offline, the local setup allows you to use DeepSeek anytime, anywhere.
Future-proof
Currently, DeepSeek’s services are free, but this is unlikely to last forever. At some point, they may need to be monetized, and usage restrictions or subscription fees may appear. With the local model, you can avoid these restrictions altogether.
Flexible
With the local version, you’re not limited by default settings.Want to fine-tune the model?
Integrate it with other tools? Build a custom interface?DeepSeek R1’s open source nature offers you endless possibilities.
Summary
At the moment, it is still unclear how DeepSeek handles user data.
If you don’t care about data privacy, using the web or mobile apps may be a better choice. They are easier to use and offer features such as DeepThink and web search.
But if you care about where your data goes, the local model is a good alternative to consider.
The DeepSeek model is designed to run well even on hardware that is not particularly powerful.
While larger models like DeepSeek-R1-Zero require a distributed GPU setup, the lite version makes it possible to run smoothly on a single GPU with lower memory requirements.
