ModelPeek
How to use ModelPeek to see which local LLM runners are running (Ollama, LM Studio, MLX, llama.cpp), which models are loaded and their memory, CPU and GPU use.
The question it answers: which local LLMs are loaded, and how much memory and GPU are they using?
What you see
One card per runner that is running:
- Ollama: the models it has in memory, with size, how much is on the GPU (“100% GPU” or a GPU/CPU split), quantization, context length and when it will unload, from Ollama’s own API.
- LM Studio: the models it has loaded, from its local server. Turn the server on in LM Studio’s Developer tab to see them; without it, ModelPeek still shows LM Studio’s memory and CPU.
- MLX, llama.cpp and llamafile: the model each was started with.
Every card shows the runner’s total memory, CPU and GPU share, and its processes.
How to use it
- Open ModelPeek when memory is short or the fans spin up.
- Find the runner with a large model still loaded.
- Unload the model or quit the runner from its own app or command (
ollama stop <model>, LM Studio’s Eject) to get the memory back.
Good to know
- Ollama and LM Studio are asked through their APIs on 127.0.0.1 (ports 11434 and 1234), and only while they run. Nothing leaves your Mac. A runner on another port says it did not answer.
- Command lines can contain API keys. ModelPeek keeps only the model name from a runner’s command line and never stores or shows the rest.
- It reads every few seconds only while its screen is open.