SolarOS 4.17.7 manual · app

zoo

Browse the hosted model zoo and install complete model bundles on mounted SD or flash storage. The text interface works on the display and VT100 port shells. It requires the Zoo package, ESP32-S3, PSRAM, and networking.

zoo
zoo refresh
zoo source [HTTPS-CATALOG-URL|reset]
zoo storage [sd|flash]
zoo install ID VERSION

The default source is https://solar-os.eu/zoo/catalog.json. The first launch fetches it; later launches use the cached catalog. r refreshes the cache. Up/Down and Page Up/Page Down select entries, Enter opens details, Alt+S searches, and d downloads the selected version. Arrows and page keys scroll long details. * marks an installed bundle and ! marks an incompatible entry. c or Esc cancels a running operation. q or Esc returns from details or exits the browser; the app-exit key exits from either view.

Source and storage preferences persist. Without an explicit storage preference, Zoo uses mounted SD, then flash. Files are stored under dl/zoo/ on that volume; each installed bundle is in models/ID/VERSION/. Downloads are streamed to storage, checked against archive, manifest, model, and asset hashes, and installed through a staged directory swap. Interrupted downloads preserve the previous installation.

The browser shows task, modality, result adapter, license, validation status, and reported free-memory requirements. Runtime compatibility is not a device inference test, and unreported memory remains unknown. The device accepts catalogs up to 512 KiB and 256 entries; larger catalogs fail without replacing the cache. Installation also needs enough storage for the ZIP and extracted files.

Installation does not load a model. Use the displayed bundle.json path with model bundle PATH or the native, Python, or Lua inference APIs. Loaded models remain resident until explicitly unloaded or the device reboots. See Model bundles for inputs, results, and execution, and Native image pipelines for a Zoo-to-camera workflow and a Python detection example. Catalog task names describe the model; check its result adapter before selecting a script. Raw tensor bundles need model-specific interpretation.