# Install mozo on CPU, CUDA or Apple silicon Mozo is one package on PyPI. It needs Python 3.10 or newer, and installs no model packages: every model's inference code is inside mozo. ```bash pip install mozo ``` That installs twelve dependencies: torch, torchvision, numpy, OpenCV (headless), Pillow, [PixelFlow](https://emrah.fr/pixelflow/), FastAPI, uvicorn, python-multipart, click, regex and ftfy. Check it worked: ```bash mozo version ``` ```text Mozo version 1.10.0 ``` ## Weights arrive on first use Installing downloads no weights. The first time you load a model, mozo downloads its weights, checks them against the sha256 recorded in the package, and caches them in `~/.cache/mozo`. The next load reads the cache. A first load can take minutes for a multi-gigabyte model; every load after that takes seconds. To download ahead of time, for example while building a container image: ```python from mozo import weights weights.fetch("rfdetr", "medium", device="cuda") ``` Pass the device the model will run on: it decides which runtime a load picks, and so which files it needs. To keep the cache somewhere else, set `MOZO_CACHE`. To refuse downloads entirely, set `MOZO_OFFLINE=1`. [Configuration](/mozo/deployment/configuration/) has both, and how to feed an offline machine from a mirror. ## Optional runtimes Every model runs on torch. Some families also publish ONNX or CoreML graphs, which run without torch's overhead on some devices. The extras install what those need: ```bash pip install 'mozo[onnx]' # onnxruntime pip install 'mozo[coreml]' # coremltools, macOS only ``` No model needs an extra. Without one, mozo simply never picks those graphs. With one installed, `runtime="auto"` may pick them where they are the faster choice: CoreML for RF-DETR, YOLOv8, YOLO12 and SAM 3 on Apple silicon. [Runtimes](/mozo/deployment/runtimes/) lists which family publishes what. ## NVIDIA GPUs Mozo uses whichever torch is installed. On Linux, the default `pip install torch` already includes CUDA. Where it does not, install the CUDA build of torch first, from [pytorch.org](https://pytorch.org/get-started/locally/), then mozo: ```bash pip install torch torchvision --index-url https://download.pytorch.org/whl/cu128 pip install mozo ``` Mozo picks the best device it finds: `cuda`, then `mps`, then `cpu`. Ask for one yourself with `device=`: ```python import mozo model = mozo.get_model("rfdetr/medium", device="cpu") ``` For ONNX on a GPU, install `onnxruntime-gpu` instead of the `onnx` extra. ## Apple silicon Torch's `mps` device runs on Apple silicon with no extra install. A few operations torch has not implemented for `mps` yet raise an error unless they are allowed to fall back to the CPU. `mozo start` allows it by itself. From Python, set it before torch is imported: ```bash export PYTORCH_ENABLE_MPS_FALLBACK=1 ``` Install the `coreml` extra for the families that publish CoreML graphs. On Apple silicon those are the fastest way to run them, and `auto` picks them by itself. ## From source ```bash git clone https://github.com/emrahnazif/mozo cd mozo pip install -e . ``` The workflow editor's built files are committed, so you need npm only to change the editor, never to install mozo.