Check your phone.
You need Android 9 or newer and an arm64-v8a device, or an x86_64 environment. Leave room for the app and the model weights, which are separate downloads. Available RAM matters as well as storage: a model that fits on disk may still be too large to load.
This development build has no recorded physical-device acceptance test. Start with a small text model and a modest context length. Larger models and longer conversations need more memory.
Download and install.
Open the download page, complete the security check and save the APK. You can download on your phone or transfer the file from a computer.
Compare the APK’s SHA-256 with the value on the download page using a trusted checksum tool. Open the APK and, if Android asks, allow your browser or file manager to install apps from this source. Keep Play Protect enabled. Turn the installation permission off again when you’re done.
The current APK is signed with an Android debug certificate. If Android reports a conflicting app signature, export your existing chats or encrypted backup before considering an uninstall: uninstalling removes local data.
Choose a starting point.
Open Fieldwork and choose Find a model for my phone or Import a model I already have. Both paths are available without signing in.
The catalogue provides model details and download requirements. Review the model’s licence, file size and estimated memory before downloading. The small SmolLM2 135M instruction model is a starting option for a basic text test; it is not a guarantee of quality or compatibility on every phone.
To import, choose a supported GGUF with Android’s file picker. For a split model, select every numbered part together. Wait for validation and copying to finish. Imported weights are stored in the app’s private files.
Try a real reply.
Go to Models → Installed → Quick test. This loads the model and generates a short answer, recording measurements on your phone. An estimated fit is different from a successful test.
Then open Chats → New conversation. Ask a question, watch the reply arrive, and use Stop whenever you like. If loading fails, try a smaller model or a shorter context. The app keeps the readable transcript when history no longer fits in the model’s context.
Bring your own context.
Open Workspace and create a project. Add project instructions and import a PDF, text, Markdown or source-code file. Select the sources available to a conversation, then open a citation to inspect the retrieved excerpt and its page or line location.
Check that a cited passage supports the answer. Language models can make mistakes, and a reference alone is not proof. Sources are scoped to their project.
Disconnect when you’re ready.
Once the model is installed, text inference works locally. Enable Offline only in Settings to stop external app transfers. You can also import weights while already offline.
Offline dictation requires Android 12+ and an installed on-device recogniser. Read-aloud needs an offline voice. Image questions require the supported MobileVLM V2 model and its exact matching projector. OCR needs the separately imported English Tesseract asset. These are optional; text chat works independently.
Keep a copy of the work that matters.
Use an encrypted backup export for supported app data, and keep its passphrase somewhere safe. Weights and original source files are excluded, so retain those separately. Markdown and JSON conversation exports are plaintext. Optional app lock restricts UI access; it does not independently encrypt the app’s database.
Understand storage and privacyCheck the Android version, architecture, free storage, model type and context size. Try a smaller model and run Quick test again. For Android security or installation warnings, review the message before proceeding.