Ollama on a 48GB MacBook Pro: What Actually Runs Well?

I tried Qwen3.6, Qwen3-Coder, and GPT-OSS 20B on my 48GB MacBook Pro with M4 Pro. Compare download sizes, launch commands, and where I'd start.

Quick Summary

What to know first

All three Ollama models felt responsive on my 48GB MacBook Pro with M4 Pro. I'd start with Qwen3.6 for general use, Qwen3-Coder for code, or GPT-OSS 20B for reasoning.

+ MacBook Pro (48GB)

I knew my MacBook Pro was powerful. But I wasn’t sure what to expect from running AI right on the laptop. Would it respond quickly enough that I’d actually use it, or would I try it once out of curiosity and forget about it?

I tried three models in Ollama on the same M4 Pro MacBook Pro I use for development. The nicest surprise was that all three felt responsive. I didn’t notice the sort of lag that would make me think twice about opening one again.

Why Use Ollama on a Mac?

New to local AI? Start with what Ollama is and how to try your first model.

  • Keep prompts and documents on your Mac. With a local model and no external tools, Ollama can process them without sending them to a cloud server. That can help with privacy requirements, but it doesn’t replace your organisation’s data rules. Ollama’s privacy FAQ explains the difference between local and cloud use.
  • Work without Wi-Fi. Once you’ve downloaded a local model, you can use it without an internet connection. Download it before you leave, not when you’re already on the train.
  • Run an AI agent with a local model. An agent can do more than answer a question: for example, it can look through a project and propose a code change. Ollama supports coding agent integrations, but check what each integration sends online before giving it sensitive files. My impressions here are about running the models, not testing a full agent workflow.

Ollama Models I Tried on a 48GB MacBook Pro

ModelTerminal commandDownload sizeWhat I'd use it for
Qwen3.6 35B-A3Bollama run qwen3.6:35b-a3b~23 GBGeneral and technical questions
Qwen3-Coder 30Bollama run qwen3-coder:30b~19 GBCode review and refactoring
GPT-OSS 20Bollama run gpt-oss:20b~14 GBReasoning through several constraints
Qwen3.6 35B-A3B
Terminal command
ollama run qwen3.6:35b-a3b
Download size
~23 GB
What I'd use it for
General and technical questions
Qwen3-Coder 30B
Terminal command
ollama run qwen3-coder:30b
Download size
~19 GB
What I'd use it for
Code review and refactoring
GPT-OSS 20B
Terminal command
ollama run gpt-oss:20b
Download size
~14 GB
What I'd use it for
Reasoning through several constraints

These are the approximate download sizes listed by Ollama for the linked tags, not the amount of RAM each model uses while running. Run ollama list to check the sizes on your Mac.

If you’re starting from scratch, follow Ollama’s macOS setup guide. Once the app is running, pick one of these commands. Allow time and disk space for the first model download; that wait is separate from how the model feels once it’s running.

Qwen3.6 35B-A3B: My Starting Point for General Use

I’d reach for Qwen3.6 when I want one model for a mix of everyday questions and technical ones. Start here if you just want to see what local AI feels like on the Mac.

ollama run qwen3.6:35b-a3b

Prompt to try: When would you pick Redis over an in-process cache for a small web app? Compare latency, invalidation, and operational overhead.

Qwen3-Coder 30B: The One I’d Try for Code

Qwen3-Coder is the one I’d pick when I already have a snippet in front of me and want a second pair of eyes. Start with a small, self-contained example so you can check the answer yourself.

ollama run qwen3-coder:30b

GPT-OSS 20B: The One I’d Try for Reasoning

I’d use GPT-OSS 20B for a question where I want the model to work through a few constraints. It also supports tool calling, but opening it in the terminal isn’t a test of a full tool workflow.

ollama run gpt-oss:20b

Prompt to try: Design a retry schedule with up to three retries in five seconds. Explain how you’d handle a response with a Retry-After header.

Done experimenting? To free disk space, remove a downloaded model with ollama rm gpt-oss:20b, replacing the tag with the one you no longer need.

Do You Need 48GB for Ollama?

No, 48GB isn’t a requirement for Ollama. I tried these three models on my 48GB M4 Pro, so I can’t say how well the same models run on a Mac with less memory. These are hands-on impressions, not a speed ranking or a test of long prompts. How much text you give a model and what else is open on your Mac can change how it feels.

On a Mac with less memory, I’d start with a smaller model or give it less text to work with. Longer context uses more memory, so the trade-off isn’t just which model you can open. It’s also how much code or conversation it can work with at once. If you’re shopping for a Mac, my MacBook Pro configuration guide explains why I’d choose memory based on the work you actually do.

My takeaway is simple: if you already have an Apple silicon MacBook, I’d try Ollama on the Mac you have. Start with one model that fits your memory and the task you have in mind. All three felt responsive on my 48GB M4 Pro, but you don’t need to start with models that large.

Ollamalocal AIMacBook Pro M4 Pro48GB unified memory