How to run an LLM on a Mac with Apple silicon

You can run an LLM on a Mac with Apple silicon using three free tools: Ollama, LM Studio or Apple's own MLX framework. On October 23, 2026 all three use the Mac's GPU, and the one number that decides which models you can run is unified memory, because the CPU and GPU share it. This guide covers the requirements, how much memory each model size needs, how to set up each tool and the memory limits that slow big models down, based on the official documentation read on October 23, 2026.

What you need

• An Apple silicon Mac (M1 or later). Ollama supports Apple M-series chips with CPU and GPU, and runs on Intel Macs on the CPU only. LM Studio does not support Intel Macs at all.
• macOS 14 Sonoma or newer. Both Ollama and LM Studio require it.
• 16 GB of memory or more. LM Studio recommends 16 GB and says 8 GB Macs can still work with smaller models and modest context sizes.
• Free disk space. Ollama warns that models can take tens to hundreds of gigabytes.

Memory decides the model

On a PC the limit is the graphics card's VRAM. On a Mac the GPU uses the same unified memory as everything else, so a 16 GB Mac has 16 GB for macOS, your apps and the model together. The download size of a model is a good first estimate of the memory its weights need, and the conversation adds more on top, because the context is kept in memory too.

Table: download sizes in Ollama's library (checked October 23, 2026) and the Mac memory we suggest

The last column is our own rule of thumb: the model plus room for macOS, your apps and the context. Smaller quantized versions use less memory at some cost in quality; our explainer on quantization covers the trade-off, and our guide on how much VRAM an LLM needs has the full calculation.

Option 1: Ollama

Ollama is the simplest choice if you are happy with a terminal. Its documentation recommends mounting ollama.dmg and dragging the app into Applications; on first start it offers to create a link so the ollama command works in Terminal. Ollama uses the Mac's GPU through Apple's Metal API. Then run a model:

ollama run qwen3:8b

Check that the model fits with ollama ps: the PROCESSOR column should read 100% GPU. If part of it says CPU, the model or its context does not fit and answers will be much slower. Models and settings live in ~/.ollama. Our list of Ollama commands covers the rest.

Context is the setting most people miss. Ollama picks a default from the available memory: 4k tokens below 24 GiB, 32k between 24 and 48 GiB and 256k from 48 GiB. It recommends at least 64,000 tokens for agents, web search and coding tools, which you set with the slider in the app's settings or with OLLAMA_CONTEXT_LENGTH=64000 ollama serve. More context needs more memory, so on a 16 GB Mac raise it carefully.

Option 2: LM Studio

LM Studio is the choice if you prefer a window to a terminal. It runs GGUF models through llama.cpp on every platform, and on Apple silicon Macs it also runs models in Apple's MLX format. You manage the engines, which LM Studio calls runtimes, with Command-Shift-R, and its lms command can download MLX versions with the --mlx flag. When a model is offered in both formats, try the MLX one first on a Mac. Our comparison of Ollama and LM Studio explains the other differences.

Option 3: MLX LM for developers

MLX is an array framework for machine learning on Apple silicon from Apple's machine learning research team, built around unified memory: arrays live in shared memory, so the CPU and GPU work on the same data without transferring it. Its mlx-lm package runs and fine-tunes language models from Python:

pip install mlx-lm
mlx_lm.chat

Without options it loads mlx-community/Llama-3.2-3B-Instruct-4bit, and the --model flag takes any of the thousands of MLX models in the mlx-community organization on Hugging Face. It is the route to choose if you want to script generation or fine-tune a model with LoRA on your Mac.

When a large model is slow

macOS limits how much memory the GPU can keep wired, that is, locked in place. The mlx-lm documentation explains that models large relative to the Mac's memory can be slow, that on macOS 15 or later it tries to wire the model's memory to speed it up, and that if it warns about the model size you can raise the limit with:

sudo sysctl iogpu.wired_limit_mb=N

N should be larger than the model in megabytes and smaller than the Mac's memory. Our advice is to leave several gigabytes for macOS and your apps, because the model and the system share the same memory.

Which one to choose

• First model, terminal is fine: Ollama. One command, and its local API is what many other apps connect to.
• Prefer a graphical app or want MLX without code: LM Studio.
• Writing Python or fine-tuning: MLX LM.

Whatever you choose, the steps for running an LLM locally apply, and if you are still choosing hardware, our comparison of NPUs and GPUs explains what each chip does.