Apple Silicon is quietly one of the best places to run a local LLM, because the CPU, GPU, and memory share one fast pool - so a MacBook with enough RAM can run models that would need an expensive discrete GPU on a PC. This guide covers the two easiest ways in, the MLX-versus-GGUF choice that is specific to Macs, and how to size your RAM. For the wider picture, see our run AI locally privacy guide and the best local models by VRAM tier.
The two easiest ways in
Ollama (command line, simplest)
Ollama is the quickest start. It needs macOS 14 Sonoma or later. Install it with:
curl -fsSL https://ollama.com/install.sh | sh
ollama run gemma4:12bThat downloads and runs a model in one command. Ollama’s default engine runs GGUF models through a llama.cpp-derived backend. As of the Ollama 0.19 preview it can also use Apple’s MLX backend to better exploit unified memory - but that path currently requires a Mac with more than 32GB of memory and, in the preview, accelerates only specific models.
LM Studio (graphical, friendlier)
LM Studio is a desktop app with a model browser and a chat UI. It is Apple Silicon only (no Intel Macs), needs macOS 14 or later, and 16GB of RAM is the sensible floor. Its advantage on a Mac is a first-class MLX engine: you can switch a model between MLX and GGUF in the UI and pick whichever runs better on your machine.
MLX vs GGUF, the Mac-specific choice
- GGUF is the cross-platform quantized format used by llama.cpp and Ollama. It runs everywhere and has the widest model selection.
- MLX is Apple’s own array framework, designed so that “arrays live in shared memory,” avoiding the CPU-to-GPU copies that cost time. On Apple Silicon it can be faster for the models that support it, and the mlx-lm tools pull ready-made models from the
mlx-communitylibrary on Hugging Face.
The practical answer: start with whichever your app defaults to (GGUF in Ollama, either in LM Studio), and if a model feels slow, try the MLX build of the same model. There is no wrong choice - it is worth a two-minute test.
How much RAM you need
Token generation on a Mac is limited by memory bandwidth more than raw compute - Apple’s own research pegs the M5 at 153GB/s, up 28% from the M4. A rough, practitioner-level rule (not an official spec, so treat it as a guide, not a guarantee): a model’s GGUF file size on disk is close to the RAM it needs, and only around 70-75% of your total unified memory is usable for weights, since macOS and the context window take the rest. In Apple’s measurements, an 8B model in 16-bit uses about 17.5GB, and a 30B mixture-of-experts model at 4-bit also fits under 18GB - so a 24GB Mac comfortably runs an 8B model, and a 32GB Mac reaches into the 30B-class.
Models that run well on a 16-32GB Mac
- gpt-oss:20b - OpenAI’s open-weight model, which the company says runs on systems with as little as 16GB of memory (about a 14GB download via Ollama). A strong first model to try.
- Gemma 4 - Google’s Apache-2.0 line sized for laptops:
gemma4:12b(~7.6GB) on 16GB Macs, and thegemma4:26bmixture-of-experts (~18GB) on 32GB. - Qwen3.6-35B-A3B - a current mixture-of-experts model (35B total, only 3B active per token) with a very long context window; its 4-bit build is around 22GB and fits a 32GB Mac, and it has a dedicated MLX tag on Ollama.
Quick start
# Install Ollama, then pick a model for your RAM:
ollama pull gemma4:12b # 16GB Macs
ollama pull gpt-oss:20b # 16-24GB
ollama pull qwen3.6:35b-a3b # 32GB
ollama run gemma4:12bPrefer a graphical app with a model browser? Install LM Studio, search for one of the models above, and download the MLX or GGUF build. For a ChatGPT-style web interface on top of Ollama, pair it with Open WebUI.
Everything here runs entirely on your Mac. Your prompts never leave the machine, there is no per-query cost, and it works offline.