Guides · Updated
Local AI on a Mac with Apple Silicon
Why M-series Macs are good at local AI, how much memory you need, which apps to use (Ollama, LM Studio, MLX), and what to check before you buy.
Macs with Apple Silicon (M1 and later) are some of the simplest machines for local AI. There are no GPU drivers to install, the popular apps use the Mac’s GPU out of the box, and the GPU can draw on most of the Mac’s memory. This guide covers what that means in practice, and where a Mac falls short of a PC with a big graphics card.
Why Macs work well
- Unified memory. The CPU and GPU share one pool of memory, so a 32 GB or 64 GB Mac can load models bigger than most PC graphics cards can hold.
- It just works. Ollama, LM Studio and llama.cpp all use Apple’s Metal for GPU acceleration automatically.
- Quiet and efficient. A Mac mini or MacBook Pro runs models without the noise and power draw of a gaming PC.
Where they fall short
- Speed. At the same model size, a Mac usually generates text more slowly than a high-end NVIDIA card, and base chips are much slower than Pro, Max and Ultra chips.
- Long prompts. Reading a long document or a big chunk of code before replying is noticeably slower on a Mac than on a strong desktop GPU.
- No upgrades. Memory is part of the chip, so you can’t add more later. Buy what you’ll need up front.
- Some tools are NVIDIA-first. Chat and speech work well, but some image, video and model-training tools are built around NVIDIA’s CUDA and run slower or not at all on a Mac.
- Intel Macs run models on the CPU only, which is slow.
How much memory do you need?
macOS keeps part of the memory for itself, so the GPU can typically use around two-thirds to three-quarters of it.
| Mac memory | Good for | Try |
|---|---|---|
| 8 GB | Trying it out with small models; tight with a browser open | qwen3.5:4b, llama3.2:3b |
| 16 GB | Everyday chat, summarising and writing with 7-12B models | qwen3.5:9b, gemma4:12b |
| 24 GB | 12-14B models with room for long chats; ~20B models are tight | qwen3:14b, gpt-oss:20b |
| 32 GB | Strong 20-27B models | qwen3.6:27b, gemma4:26b |
| 64 GB | Larger mixture-of-experts models; 70B is tight | qwen3.6:35b, llama3.3:70b |
| 96-128 GB | The biggest open models that run at home | gpt-oss:120b |
For your exact Mac, choose Apple Silicon Mac in the GPU picker. Install steps and where files live are in How to install Ollama on a Mac.
Buying a Mac for local AI
- Memory first. It decides which models you can run at all. 16 GB is a sensible minimum; 32 GB or more opens up the strongest models in our guides.
- Then the chip tier. Speed tracks memory bandwidth, and Pro, Max and Ultra chips have much more of it than base chips. An older Max-chip Mac with plenty of memory can be a better local AI machine than a newer base model.
- Then the body. The MacBook Air has no fan and slows down during long jobs. A MacBook Pro, Mac mini or Mac Studio can sustain full speed.
See our Mac mini and Mac Studio picks, which compares them with other unified-memory machines. How to choose between a Mac mini and other small boxes: Choosing a mini PC for local AI.
Which app to use
| App | Best for |
|---|---|
| Ollama | A background service that coding assistants, chat apps and scripts can connect to |
| LM Studio | A desktop app that shows whether a model fits before you download it, with a mature MLX engine |
MLX (mlx-lm) | Apple’s own machine-learning framework, used from the command line, for tinkerers |
| llama.cpp | The engine behind many apps, with Metal support built in (brew install llama.cpp) |
GGUF or MLX? GGUF is the format Ollama and llama.cpp use on every platform. MLX models are built for Apple Silicon and often run a little faster on a Mac. LM Studio lets you pick either. To try MLX directly:
pip install mlx-lm
mlx_lm.generate --model <model-name> --prompt "Hello"
Ready-converted MLX models are published on Hugging Face, many under the mlx-community account. Comparing the two main apps: LM Studio vs Ollama.
Check it’s running well
- On the GPU? Run
ollama ps. The PROCESSOR column should read100% GPU. - Enough memory? Open Activity Monitor → Memory and watch Memory Pressure. If it turns yellow or red, macOS is swapping to the SSD and replies will crawl. Use a smaller model or close other apps.
- GPU busy? In Activity Monitor, Window → GPU History shows how hard the GPU is working.
Raising the GPU memory limit (advanced)
On recent versions of macOS you can let the GPU use more memory with sudo sysctl iogpu.wired_limit_mb=<megabytes>. The setting resets when you restart. Leave several gigabytes for macOS, because setting it too high can make the whole Mac freeze. Most people don’t need this; picking a model that fits is the safer fix.
Mac tips
- Plug in for long jobs. Low Power Mode slows generation down.
- Store models elsewhere. Models take up a lot of disk space. To keep them on an external SSD, set
OLLAMA_MODELSwithlaunchctl setenv(see the Mac install guide), then restart Ollama. More on freeing space: Keeping local models up to date and tidy. - Keep apps updated. Ollama, LM Studio and MLX regularly improve Apple Silicon performance.
More than chat
- Speech: whisper.cpp uses Metal and runs well on any M-series Mac. See Local speech-to-text and text-to-speech.
- Images: Draw Things is a free native app, and ComfyUI has a desktop app for Apple Silicon. Expect slower results than a strong NVIDIA card. See Generating images locally.
- Coding: connect a local model to your editor with the local coding assistant guide.
- Fine-tuning: MLX can train LoRA adapters on a Mac. See Fine-tuning a small model with LoRA.
- A nicer chat window: see Open WebUI and alternatives.
On a MacBook? The laptop-specific gotchas are in Running local AI on a laptop. If ollama ps shows the CPU, or a model will not load, see When it doesn’t work.
Sources: Ollama macOS docs, LM Studio docs, MLX on GitHub.
Get the Local AI weekly
New models worth running, what fits which GPU, and setup fixes. One short email a week, free. Unsubscribe any time.
Local-LLM App Starter
A tested starter for building your own app on Ollama: chat, RAG and a GPU-fit checker. Skip the setup hours.
kits.lilaforge.com → VibeEngineBest local LLM tools, ranked
Ollama, LM Studio, Open WebUI and llama.cpp scored from public evidence, with top picks.
vibeengine.lilaforge.com →