What size local model should I start with on a Mac with 48 GB of RAM?

Running a model on your own Mac gives you privacy and avoids per-question costs. But choosing one that’s too large makes everything slow, and a memory estimate can help you choose the right size the first time.

Parameters are the model's “learned numbers”; the more parameters it has, the more memory it uses. A rough rule of thumb is to reserve about half your RAM for the model—around 24 GB here. It's a conservative starting point (your Mac can use more), not a technical limit.

Here's the math: in the most common compressed version (quantized to 4 bits), each billion parameters takes around 0.6 GB. A 20-billion-parameter model comes to about 12 GB, and the rest of the 24 GB covers the conversation's memory. The system and other programs continue to use the other half.

Don't start by choosing the biggest model that fits on the machine: with some headroom, the computer remains usable while the model runs. On a Mac, look for the MLX version of the model, optimized for Apple's chips.

Unified memory on Mac

On Macs with Apple chips, memory is unified: the processor and graphics use the same RAM. That means the 48 GB can run the model without a separate graphics card.

In return, everything shares the same space: the model, browser, open apps, and the system itself.

Model size and conversation length

The model uses memory in two ways. First, there’s the model file itself: at 4 bits, about 0.6 GB per billion parameters. Quantization (compressing the model’s numbers) at 8 bits uses close to twice as much, as an estimate.

Second, there’s the context (how much conversation and how many documents the model considers at once). The more text you paste in, the more memory it uses, in addition to the model file.

The calculation, step by step

Set aside about half the RAM: 24 GB. A 20-billion-parameter model at 4 bits uses about 12 GB. That leaves about 12 GB for context, enough for long conversations and medium-sized documents without strain.

A much larger model may load, but it leaves little room to spare and can slow down the Mac for other work.

Tools to get started

LM Studio and Ollama download and run models in a few clicks or with a command. On Hugging Face (a public model repository), look for MLX versions (a format optimized for Apple chips) or 4-bit quantized versions. Test the model on your real tasks before deciding it works for you.

1. Check how much memory your Mac uses during your normal work in Activity Monitor.
2. Install LM Studio or Ollama.
3. Download a model with 15 to 20 billion parameters at 4 bits, preferably an MLX version.
4. Test it on three or four real tasks and note the quality and speed.
5. Watch memory pressure while using it with your usual apps open.
6. Move up in size only if you have room and the current quality isn’t enough.

• Choosing the largest model that fits: there’s little memory left and the Mac gets slow.
• Forgetting to account for context: long documents use memory beyond the model file.
• Judging a model by a single test: always compare it on your real tasks.

• 48 GB ÷ 2 ≈ 24 GB for the model.
• 4 bits: ~0.6 GB per billion → 20 billion ≈ 12 GB.
• On a Mac, use the MLX version.

• CourseKit Local — the local labSimulate how much model will fit in your memory before downloading.
• Projectjarvis-localA voice assistant that runs 100% on your machine, with Ollama.
• CourseCCOpen — Claude Code for free or next to nothingUse a local model on your machine to power the agent.

Some of these materials are in Portuguese.