How I Set Up DeepSeek R1 Locally on an M2 MacBook

How I Set Up DeepSeek R1 Locally on an M2 MacBook

By Ebuka Onah

Running frontier-grade AI models locally used to sound ridiculous. Today, it's surprisingly practical—especially on Apple Silicon.

I recently installed DeepSeek R1 on my M2 MacBook, and honestly, the experience was far smoother than I expected.

Local AI is no longer a hobbyist experiment. It's becoming a legitimate workflow.

Why Run DeepSeek Locally?

Three reasons:

  • Complete privacy
  • No API costs
  • Instant offline access

Your prompts never leave your machine. That's a huge advantage for sensitive projects.

What You'll Need

  • Apple MacBook with M2 chip
  • At least 16GB unified memory
  • macOS Sonoma or newer
  • About 20GB of free storage
  • Ollama installed

If you're working with 8GB RAM, things get tight very quickly. Physics remains undefeated.

Installing Ollama

Ollama makes local model deployment almost embarrassingly easy. 

curl -fsSL https://ollama.com/install.sh | sh

After installation, verify everything is working: 

ollama --version

Downloading DeepSeek R1

Pull the model directly from Ollama's registry: 

ollama run deepseek-r1

The first download can take a while depending on your connection. Large models are not known for their modesty.

The 7B version offers the best balance between speed, quality, and MacBook sanity.

Performance on an M2 MacBook

  • Response speed: 20–35 tokens per second
  • RAM usage: 8–12GB
  • Thermals: Warm, but manageable
  • Fan noise: Minimal on most workloads

Apple's unified memory architecture makes an enormous difference here.

My Recommended Setup

For daily use, I pair Ollama with Open WebUI. 

docker run -d \
-p 3000:8080 \
--add-host=host.docker.internal:host-gateway \
-v open-webui:/app/backend/data \
ghcr.io/open-webui/open-webui:main

This gives you a polished ChatGPT-like interface running entirely on your machine.

What Broke First

A few things.

  • Large context windows quickly consumed memory
  • Multi-model switching caused swapping
  • Long reasoning chains heated the machine noticeably
  • Background apps became your enemy

Close Chrome. Yes, all 47 tabs.

When Local Beats Cloud

  • Private coding sessions
  • Offline writing workflows
  • Rapid experimentation
  • Low-latency iteration
  • Cost-sensitive usage
For many everyday tasks, local models are already "good enough"—and good enough scales quickly.

Final Thoughts

DeepSeek R1 running locally feels like a glimpse of where personal computing is heading.

AI is shifting from something you access online to something you own.

And once you experience that, cloud-only workflows start feeling strangely old-fashioned.