Stop Running the Biggest Model Your 128GB Mac Can Hold

I made a video on this after starting with what felt like an obvious question: on a 128GB Mac, is the best local model just the biggest one that fits? The instinct is to load the largest model and call it done. The people actually doing local AI on this hardware say that is the wrong move, and their answers changed how I would set it up. Here is the whole map so you do not have to watch to use it.

What close to 100 replies agreed on

Before recording I put the question to local LLM users on Reddit and got almost 100 answers. The consensus was a workflow, not a winner. You keep a few models around and pick per task.

For daily coding and action loops, 27B to 35B models win. Qwen 3.6 35B, Qwen Coder 30B, and Gemma 4 31B are the ones people keep running in the background. Their speed is what makes the tight edit-run-edit cycle bearable, and they leave enough memory to keep VS Code, Docker, and browser tabs open. You feel the lag from a giant model most in exactly this loop.

Big models earn their keep in planning

The larger models are not useless, they just belong in a different slot. Qwen 122B-A10B is a sparse mixture-of-experts model, meaning only about 10 billion of its 122 billion parameters activate per token, so it can plan a bigger task without running at full-model cost. On my machine Ollama shows it as a 4-bit build with a 262K context window, about 81GB on disk. Let it draft the plan, then hand execution to a fast 27B model.

The edge case and the real advantage

DeepSeek V4 Flash through ds4 and Dwarfstar sits at the boundary of what is practical here. Treat it as a case you reach for when the extra wait is worth it, not a daily driver.

The reason all of this works on a Mac is unified memory. It changes the workflow by letting a planner and an executor coexist and swap cleanly. A GPU box can beat it on raw speed, but for this mixed setup the Mac is usually the easier machine to live with.

Watch the full breakdown in the video above, and check the Reddit thread linked in the description for the raw answers. For more on running models locally, openclawdc.com has the related guides.

I specialize in helping teams design local model workflows that fit their hardware and their work. Book a call at cloudyeti.io/meet.