Model download
Cove's models live on your device. They're downloaded once on first launch, shared across all four apps, and never sent back.
The Gemma 4 E2B weights are hosted on Hugging Face (the LiteRT-LM community mirror of Google's official Gemma release). The download is direct, with TLS 1.2+. Hash verification confirms file integrity before loading. No third-party CDN, no Cove server in the chain.
Gemma 4 ships in two variants: E2B (default, ~2.5 GB, 2-billion effective parameters) covers everyday use; E4B (optional upgrade, ~3.6 GB, 4-billion effective parameters) gives higher accuracy on long technical text and rare languages. Both are Q4_0 quantized for efficient on-device inference. Free and Pro use the same file — upgrading to Pro never re-downloads the model, it only removes the 5-runs-per-day cap.
Primary location: /sdcard/GemmaApp/models/ (shared external storage). This survives app reinstall — if you uninstall Travel and install Voice, the model is reused. Fallback: app-private directory if external storage is unavailable. To check, navigate to /sdcard/GemmaApp/models/ in any file manager.
Travelling somewhere with no Wi-Fi? Pre-download the Gemma 4 E2B weights on your home connection and copy the .litertlm file into /sdcard/GemmaApp/models/. On next launch, Cove detects the file and skips the download. Useful for shared family phones or pre-trip prep.
All four Cove apps look at the same /sdcard/GemmaApp/models/ directory. Installing more Cove apps does not duplicate the model. If you uninstall all Cove apps and want to free disk space, manually delete the model file.
Cove checks the model file's SHA-256 hash before loading. If verification fails, the app will re-download or prompt you to delete and retry. We do not load unverified weights — this is part of the security commitment.
