DeepSeek is an open source artificial intelligence developed in China, which stands out for its advanced reasoning capabilities in complex tasks (such as mathematics and coding).
Thanks to its flexible design and the possibility of running its models locally, DeepSeek is presented as an affordable alternative to other large language models, without depending on cloud services.
This article provides an accessible overview of how to download, install, and use DeepSeek locally, as well as clarifying important technical concepts so you can take full advantage of its capabilities. But it is not a detailed guide to each procedure.
How to download and run Deepseek locally
There are two main methods to use DeepSeek on your computer:
Command Line Tools (like Ollama)
With Ollama, you can download and install various variants of DeepSeek using simple commands.
This option is ideal for users with basic terminal knowledge who want greater control over the inference process.
Environments with Graphical Interface (such as LM Studio)
LM Studio allows you to search, download and manage models through an intuitive graphical interface, making the task easier for those who prefer to avoid the command line.
It is especially useful for developers and enthusiasts who want to experiment with DeepSeek visually.
Hardware and Software Requirements
To run DeepSeek locally, it is important to have the right hardware:
- RAM Memory: Minimum 16 GB (although some distilled variants can operate with less).
- Processor: A modern multi-core CPU (e.g. Intel i5 or AMD Ryzen 5).
- GPU: Although there are models that can run on CPU, it is recommended to have a dedicated GPU for larger variants (for example, 7B parameter models require at least 8 GB of VRAM).
- Operating system: Compatible with Linux, macOS and Windows (using WSL2 for Linux environments).
- Dependencies: Python 3.7 or higher, along with package managers such as pip or Conda to install necessary libraries.
- Optional: Docker, to run DeepSeek in an isolated environment.
Available models and advantages
DeepSeek offers several variants adapted to different hardware capabilities:
Complete model
With very large sizes (up to 671B parameters) and designed for environments with high-performance hardware.
Distilled Versions
These are reduced and optimized versions of the full model. For example, there are variants with 1.5B, 7B or 14B parameters.
Distillation consists of training a smaller model (the “student”) using the outputs and behaviors of a large model (the “master”).
This allows it to retain much of the performance of the original model, but with lower memory and resource consumption, making it easier to run on consumer computers or mobile devices.
To further reduce the model size and optimize the inference speed, the quantization technique is applied. This involves representing the model weights using fewer bits (for example, 8 bit or 4 bit) instead of the standard 32 or 16 bit representation.
Both distillation and quantization techniques reduce the space required to store the model and allow the model to be run on devices with limited hardware. In return there may be a slight decrease in the accuracy of the calculations and answers.
Limitations and considerations
While running DeepSeek locally offers benefits like increased privacy and customization, it’s also important to keep in mind:
- More powerful versions require advanced GPUs and, in some cases, more RAM.
- Running the model on CPUs or modest GPUs may result in slower response times.
- Installation and configuration (especially using command line tools) may require basic technical knowledge.
Being a model of Chinese origin, some variants may present limitations regarding sensitive content. Running the model locally allows, in certain cases, to modify or adjust these parameters.
Frequently Asked Questions and Answers (FAQ)
What are distilled versions of a model?
Distilled versions are reduced, optimized models that have been trained from a larger, more complex model.
During the distillation process, knowledge from the original model (the “master”) is transferred to a smaller model (the “student”), which allows maintaining good performance in specific tasks, but with lower consumption of memory and computational resources.
This is especially useful for running AI on devices with hardware limitations.
What does it mean to use 8-bit or 4-bit formats?
It refers to the quantization of the model parameters. Instead of storing values in higher precision formats (such as 32 or 16 bits), 8 or 4 bits are used for each value.
This reduction reduces the size of the model and improves computational efficiency, allowing the model to run on less powerful hardware, although it may lead to a slight loss of precision in the calculations.
How can I install DeepSeek locally?
You can install DeepSeek using tools like Ollama or LM Studio. With Ollama, for example, model variants are downloaded and executed using commands in the terminal, while LM Studio offers a graphical interface that makes model management easier.
It is important to follow the recommended hardware and software requirements to obtain the best performance.
What hardware requirements do I need to run DeepSeek?
In general, we recommend a minimum of 16 GB of RAM, a modern processor (e.g. Intel i5 or AMD Ryzen 5) and, for larger variants (such as the 7B parameter), a dedicated GPU with at least 8 GB of VRAM.
It is also possible to run smaller models on CPU, although with lower inference speed.
What are the advantages of running DeepSeek locally?
Running DeepSeek locally allows you to have greater control over your data, improve privacy (as no data is sent to external servers), and customize the model to your needs.
In addition, you can save costs associated with cloud services and adapt the use of the model to environments with limited resources.
This post is also available in: