How to run a local LLM on the Apple M4 (16GB)
A step-by-step setup for beginners. With 16 GB of unified memory, the best everyday model for this Mac is Qwen3 8B, which runs faster than you can read.
What to run on the Apple M4 (16GB)
The best model for each job that fits fully in memory at the default Q4_K_M download and runs at a comfortable reading speed.
| Job | Model | Speed | Ollama command |
|---|---|---|---|
| Chat and everyday questions |  Qwen3 8B | Faster than you can read | ollama run qwen3:8b |
| Help with code | Qwen3 8B | Faster than you can read | ollama run qwen3:8b |
| Writing and editing |  Llama 3.1 8B | Faster than you can read | ollama run llama3.1:8b |
| Questions about my documents | Qwen3 8B | Faster than you can read | ollama run qwen3:8b |
| Maths, logic and homework | Qwen3 8B | Faster than you can read | ollama run qwen3:8b |
Set up Qwen3 8B in about ten minutes
Pick one app. LM Studio is easiest if you've never used a terminal.
LM Studio on Mac
- Go to lmstudio.ai and click Download for Mac.
- Open the downloaded file and drag LM Studio into your Applications folder, then open it from there.
- If it offers to download a first model, you can skip that. Click the search (magnifying glass) icon in the left sidebar to open Discover.
- Type Qwen3 8B in the search box. On a Mac you may see both MLX and GGUF versions: either works, and MLX is often a little faster.
- Choose the 4-bit (or Q4_K_M) version if a list appears, then click Download.
- Open the Chat tab, click Select a model to load at the top, pick Qwen3 8B, and type your first message.
Ollama on Mac
- Go to ollama.com/download and download Ollama for macOS.
- Open the download, move Ollama to your Applications folder, and open it. Allow it to install its command-line tool when asked.
- Open Terminal (press Cmd + Space and type "Terminal") and run the command below. The first time, it downloads Qwen3 8B, which can take a while.
ollama run qwen3:8b- When you see
>>>, type a message and press Enter. Type/byeto quit. Ollama's own app window can also chat with any model you have downloaded.
If something goes wrong
The four problems beginners hit most.
Out of memory or won't load
The model plus its working memory doesn't fit in your graphics memory. Here is how to make it fit, from the easiest fix to the most advanced.
It's really slow
Most slowness comes from the model not fitting fully in graphics memory, or from "thinking" models reasoning before they answer.
It's not using my GPU
If your graphics card sits idle while the CPU does all the work, one of these is usually the reason.
Weird, wrong or repeating answers
Local models can be excellent, but small or badly configured ones make mistakes that paid assistants don't. Here is what to check.
Run it in the cloud
Want a bigger model than the Apple M4 (16GB) can hold?
Rent a GPU with more memory by the hour, or see whether upgrading or renting is cheaper for you.
Frequently asked questions
What is the best local LLM for the Apple M4 (16GB)?
For everyday chat, Qwen3 8B: it fits in the Apple M4 (16GB)'s 16 GB at Q4_K_M and runs at about 15 tokens per second.
Should I use LM Studio or Ollama on the Apple M4 (16GB)?
Either works. LM Studio is a point-and-click app and the easiest start; Ollama is a one-line command that other apps and coding tools can connect to.
Can the Apple M4 (16GB) run a 70B model?
Not fully in memory: a 70B model needs about 46.1 GB at Q4_K_M. It would spill into system RAM and run very slowly, so renting a GPU is the practical option.