Download flow

1. Pick a model in Models. Each row shows runtime, 4-bit size, nominal / usable context and minimum RAM; a star marks the default (Qwen3-1.7B-4bit, 968 MB).
2. Headroom check. Before anything is fetched the app compares the model's size + KV cache at the default context against os_proc_available_memory() and the free disk (volumeAvailableCapacityForImportantUsage). Not enough → a sheet explains exactly what is short.
3. Download straight from huggingface.co — only the files in that repo (*.safetensors, config.json, tokenizer*.json). Progress is size-weighted across files. A gated repo without a token shows Needs a Hugging Face token with access and links to Settings.
4. Resume. Leave the app and the download pauses (foreground session, ~30 s grace); come back and it resumes with Range: from the partial file. Partial files are kept until you delete the model.
5. Loaded. Tap a downloaded model to load it into memory; the row shows Loaded and the Chat tab starts using it. Loading another model unloads the first.
6. Delete removes the snapshot folder. Clear cache in Settings removes all of them.

Where files live

Application Support/loopl/models/<org>--<name>/ — never Library/Caches (purgeable), never Documents (shows in Files and backups). The folder is marked excluded from backup so iCloud does not try to upload gigabytes.

Apple Foundation Models

Has no download step at all. If the OS is still fetching the system model you see Apple is still downloading the model; that is the OS, not the app.