# Configure the cache, downloads and served models Mozo needs no configuration: weights download on first use into `~/.cache/mozo`, and the server offers every model. These settings change that. All are environment variables, and the weight settings can also be set in code. ## Environment variables | Variable | Default | Description | |---|---|---| | `MOZO_CACHE` | `~/.cache/mozo` | Where downloaded weights are kept. | | `MOZO_OFFLINE` | unset | `1` refuses every download. A missing file raises an error naming its path, URL and sha256, so it can be placed by hand. | | `MOZO_BASE_URL` | mozo's own server | Download weights from a mirror instead. A `file://` URL to a copy of the weights tree works. | | `MOZO_ENABLE` | every model | Which models the server offers: families (`clip`) and single variants (`clip/base`), comma-separated. | | `MOZO_PREPARED_MB` | `512` | Memory the server may use for images kept by `/prepare`. The least recently used goes first. | | `PYTORCH_ENABLE_MPS_FALLBACK` | `1` under `mozo start` | Torch's own: on Apple silicon, run operations `mps` lacks on the CPU instead of failing. | `MOZO_ENABLE` and `MOZO_PREPARED_MB` are read by the server once, when it starts. ## Set it in code An application that embeds mozo can set the weight settings itself. Each wins over its environment variable, and applies to the next download: ```python import mozo mozo.configure(cache="~/MyApp/weights", offline=False) ``` `mozo.configure(cache=None)` goes back to the environment variable. The same module answers what the cache holds, without importing torch or touching the network: ```python from mozo import weights weights.is_cached("sam2", "large", device="mps") # would a load need the network? weights.download_size("sam2", "large", device="mps") # bytes a load would still download weights.cached() # every file: model, key, size, state weights.remove("sam3", "sam3") # delete it, returns the bytes freed ``` [Weights API](/mozo/reference/weights/) has every function. ## Choose what a server offers `MOZO_ENABLE` is an allow-list. Unset, the server offers all 82 models. Set, it offers only what it names: ```bash MOZO_ENABLE=rfdetr,clip/base,granite_embedding_r2 mozo start ``` A family name offers all its variants; `family/variant` offers one. A name that matches nothing is logged and ignored, never fatal: the setting can only take models away. A request for a model the server does not offer answers `404`, and `GET /models` lists only what it offers. It is an allow-list rather than a list of exclusions because the two fail in opposite directions when you upgrade. A list of exclusions would start serving whatever family the upgrade added; an allow-list serves nothing it was not told to. [Licences](/mozo/deployment/weights-and-licences/) has the list that offers only permissively licensed weights. Workflows are held to the same list. A workflow sent to `/workflow/run` that names a model the server does not offer fails at that node, with the same reason `/predict` gives, before any weights are downloaded. A node that runs a checkpoint of your own is not refused: the list decides which of mozo's published weights a server hands out. `MOZO_ENABLE` is a server setting. It does not affect `mozo.get_model`; importing mozo is not a deployment. ## Run without internet access Download on a connected machine, then carry the cache across: ```python from mozo import weights for model in ["rfdetr/medium", "sam2/large"]: family, variant = model.split("/") weights.fetch(family, variant, device="cuda") ``` Copy the cache folder to the same path on the offline machine, or point `MOZO_CACHE` at it, and set `MOZO_OFFLINE=1` so a missing file fails at once instead of trying the network: ```bash MOZO_CACHE=/data/mozo-cache MOZO_OFFLINE=1 mozo start ``` To serve weights to several machines from one place, put a copy of the weights tree on a file share or an internal web server and point `MOZO_BASE_URL` at it, such as `file:///mnt/weights` or `https://weights.internal/mozo`. The tree mirrors mozo's own: one folder per family, variant and revision, holding the files the manifest names, such as `rfdetr/medium/2026-08-18/torch-fp32.pth`. Every file is still checked against the sha256 in the package, so a mirror cannot serve different bytes.