- In this machine learning and AI tutorial, we explain how to install and run the DeepSeek-R1 model locally on a Windows computer.
- DeepSeek-R1 belongs to the class of reasoning models. Reasoning models perform better at complex reasoning problems and tasks compared to the classical large language models. Complex reasoning problems are problems that appear in mathematics, science, and coding.
- According to the information given on the GitHub page of DeepSeek, the performance of DeepSeek-R1 is comparable to the performance of OpenAI-o1. However, DeepSeek-R1 is released under the MIT license, which is a very liberal license allowing even for commercial use.
Distilled Models:
– To run the full DeepSeek-R1 model locally, you need more than 400GB of disk space and a significant amount of CPU, GPU, and RAM resources! This might be prohibitive on the consumer-level hardware.
– However, DeepSeek has shown that it is possible to reduce the size of the original DeepSeek-R1 model, and at the same time preserve the performance (of course, not completely). Consequently, DeepSeek has released a number of compressed (distilled) models whose size varies from 1.5-70B parameters. To install these models you will need from 1 to 40 GB of disk space.
– In this tutorial, we will explain how to install and run distilled models of DeepSeek-R1.
The YouTube tutorial is given below.
Installation Instructions
First download and install Ollama from the official Ollama website:
After that, verify that Ollama is installed, by opening a command prompt and type “ollama”. As the response, you should obtain the response shown in the figure below.
The next step is to install the model. Go to the Ollama website, and search for DeepSeek-R1\
Then, search for the distilled model. In this video tutorial, we will install the 14B distilled model. Click on this model, and an installation command will be generated.
The generated command is
ollama run deepseek-r1:14bExecute this command in the command prompt and the DeepSeek-R1 model will be installed and executed.