As the Tech Lead at AirConsole, I constantly strive to improve our development workflows and ensure the security of our proprietary code. Recently, I undertook a project with two main objectives:

Eat Our Own Dogfood: Test and validate AirConsole’s tooling by creating a mini-game to validate our game-dev tooling.
Maintain Code Privacy: Utilize a local Large Language Model (LLM) to assist in development without exposing sensitive automotive code to public LLMs.

To achieve these goals, I set out to run DeepSeek R1 locally on my Mac.

Step-by-Step Setup

The pipeline is simple: PyCharm → CodeGPT Plugin → Ollama → DeepSeek R1

Install Ollama on Mac

Actions:

Download: Visit the Ollama website and download the latest macOS version.
Install: Run the installer and follow the on-screen instructions, granting necessary permissions.

Notes:

Ollama serves as the intermediary between CodeGPT and DeepSeek R1, enabling local AI processing.

Download DeepSeek R1

In a terminal run and wait:

ollama run deepseek-r1

Upgrade PyCharm (or any JetBrains IDE)

That took me the longest because I couldn't figure out why I had an outdated version of CodeGPT and then a random issue on the repo made me realize that it might be the issue.

Install CodeGPT Plugin

Within PyCharm (or any other IDE)

Navigate to Preferences > Plugins.
Search for CodeGPT.
Ensure installation of version (at least) 2.15.2-241.1.
Install and restart PyCharm when prompted.

Configure CodeGPT Plugin

Start playing

To convince yourself that nothing is shared over the wire:

Disconnect internet
Activity Monitor > Window > GPU History
Select some code and start chatting with it (right side of the IDE)

Look at the GPU go mad

Enjoy living in 2025!

Results

Performance:
Running DeepSeek R1 locally is noticeably slower compared to cloud-based solutions.

Here is a real-time video on a Mac Mini M1 (16GB Ram)

Resource Management:
DeepSeek R1 is GPU-intensive. I will update to a Mac Mini M4 soon.

DeepSeek
I don’t have extensive experience with DeepSeek yet, but I find it reasonably effective for coding. However, the speed issue makes it somewhat impractical, as the slower iteration time significantly lags behind cloud-based models. I see its value in scenarios like being on a train or plane with poor internet connectivity, but it falls short in high-performance contexts where rapid code delivery is critical.

Privacy
However, in an automotive environment, where I also work, it could become a very compelling option. This is especially true when dealing with code that cannot be exposed to public LLMs, as the local processing ensures enhanced security and privacy.

Update

The author of CodeGPT (Carl-Robert Linnupuu) contacted me and told me that he had cooked with Deepseek in the last days:

He integrated DeepSeek R1 directly in the local server so that you actually do not even need ollama. Might be useful for corporate colleagues that struggle getting anything installed on their machine (diff here).

You can now choose Deepseek R1 directly from the IDE

Additionally, the user interface has been slightly refined—kudos to him for continually enhancing the design. For all these reasons, we encourage you to follow him and support his ongoing efforts to further develop this remarkable project.