{"id":171,"date":"2026-04-09T03:40:08","date_gmt":"2026-04-09T03:40:08","guid":{"rendered":"https:\/\/beginnerprojects.com\/cms\/?p=171"},"modified":"2026-04-14T03:50:28","modified_gmt":"2026-04-14T03:50:28","slug":"local-ai-running-pro-models-on-a-baseline-mac-studio-36gb-ram","status":"publish","type":"post","link":"https:\/\/beginnerprojects.com\/cms\/local-ai-running-pro-models-on-a-baseline-mac-studio-36gb-ram\/","title":{"rendered":"Local AI: Running Pro Models on a Baseline Mac Studio 36GB RAM"},"content":{"rendered":"\n

If you spend any time in the AI space, you’ve probably developed a bit of “RAM envy.”<\/strong><\/p>\n\n\n\n

You see the YouTubers and the power-users showcasing massive models running on Mac Studios with 128GB or 256GB of unified memory. It\u2019s impressive, but for most of us, it’s irrelevant. Most of us aren’t dropping $10,000 on a workstation.<\/p>\n\n\n\n

I wanted to see what was actually possible on a baseline machine. I have a Mac Studio M4 Max with 36GB of memory, and I’ve spent the last few months proving that you don’t need a super-computer to develop professional software locally.<\/p>\n\n\n\n

My “Headless” Strategy: The Best of Both Worlds<\/h3>\n\n\n\n

Here is a secret about my workflow: I don\u2019t actually “work” on my Mac.<\/strong><\/p>\n\n\n\n

I\u2019ll be honest\u2014I\u2019ve never fully adapted to macOS. The shortcuts and the window-activation workflow (having to click a window just to paste code) always felt like a friction point for me. I prefer the speed and fluidity of Linux.<\/p>\n\n\n\n

So, I did something a bit unconventional: I didn’t even install Homebrew on the Mac.<\/p>\n\n\n\n

Instead, I treat my Mac Studio as a dedicated AI Server.<\/strong> I use a basic Dell PC running Linux as my primary interface, accessing the Mac’s power via Ollama. This allows me to keep the Linux workflow I love while leveraging the incredible unified memory of the M4 Max in the background. It’s a headless powerhouse.<\/p>\n\n\n\n

The Rotation: What Actually Works on 36GB?<\/h3>\n\n\n\n

Since the fall of 2025, I\u2019ve cycled through dozens of LLMs. But as of today, my “daily drivers” for programming and web design have narrowed down to two: Gemma 4:31b<\/strong> and Qwen 3.5:27b<\/strong>.<\/p>\n\n\n\n

For a long time, qwen3-coder<\/code> was the gold standard. But the latest Qwen 3.5<\/strong> release is such a jump in capability that it has effectively replaced everything else for my coding needs. Whether I’m grinding through complex backend logic or polishing a front-end design, these two models handle it all without breaking a sweat.<\/p>\n\n\n\n

Why I Switched to Ollama (The Stability Factor)<\/h3>\n\n\n\n

When you’re working with limited RAM, stability is everything.<\/p>\n\n\n\n

I used to use LMStudio, but I kept hitting a wall: memory spikes. I’d watch my system monitor climb to 33GB, and then\u2014crash<\/em>. The whole model would drop, and I’d have to restart.<\/p>\n\n\n\n

Switching to Ollama<\/strong><\/a> changed the game. By using a precise config.yaml<\/code> via the Continue extension, my memory usage hovers steadily around 29GB. It is rock solid. I can leave a massive model like gemma4:31b running for hours on end without a single crash.<\/p>\n\n\n\n

\"Screenshot
Real-world memory consumption of a local AI model running on a baseline Mac Studio M4 Max 36GB RAM<\/figcaption><\/figure>\n\n\n\n

The Ecosystem in Action<\/h3>\n\n\n\n

To give you a real-world example: every piece of software I’ve developed for Beginner Projects<\/em> was built using this hybrid local setup:<\/p>\n\n\n\n