The time to discover a missing model download is before you board. Offline AI works best when the app, models, and files are already on your device and you have tested the task without a connection.
OGAD (Off Grid AI Desktop) can run local AI on your computer after setup. Download the resources for the work you plan to do, save the source files, and perform a disconnected test. Start with one useful task rather than preparing every available feature.
| Download OGAD | Desktop releases |
What would you like to do with Off Grid AI?
Have a feature or use case you would like us to support? Tell us what you want to do and which device you use.
Write to support@offgridmobileai.co, join our Slack community, or talk to us on Reddit.
A flight can be a useful time to draft, review notes, or work through documents. Those tasks need a local model and local material; they do not need live search or access to a server at home.
What do you actually need offline?
Choose the work first, then prepare the required models and files. A text model may be enough for writing from notes. Speech, image analysis, and image generation need their own compatible models and runtimes.
Suppose you want to review a client brief and prepare discovery questions during the flight. You need a readable copy of the brief, a local text model, and enough available memory to process the request.
| Planned task | Resources to prepare |
|---|---|
| Draft from notes | Local text model and saved notes |
| Ask about a document | Local text model and readable local file |
| Query a project collection | Local text model, indexing resources, indexed sources |
| Analyse a photo | Compatible local vision model and saved image |
| Transcribe audio | Local speech model, runtime, and saved recording |
Downloading one model does not automatically make every row ready. Prepare only the tasks you expect to use.
Which desktop version should you install?
Use the build for your supported operating system. The free core app supports Mac and Windows, while 0.0.54-beta.108 also provides Linux packages as a beta.
Complete installation while you still have a reliable connection and time to resolve a problem. Avoid making your first trial of a new model or optional runtime the last task before departure.
This guide focuses on free core chat and files. If your planned workflow needs Pro, complete its activation and check access separately. Local inference does not by itself establish that every licensed feature will remain available indefinitely without a connection.
How do you prepare the models?
Open Models and download the local model for your first task. Read its size and fit information, leave storage for the installed resources, and try a small request after selecting it.
- Choose a local text model in Models > Text.
- Wait for the download to finish.
- Select the model and send a short prompt.
- Check that the response is useful and the computer remains responsive.
- Prepare other modalities only if you need them for the trip.
Keep the selected model local. A saved configuration for a remote server does not make that server available without network access. Private addresses and remote-access tools still need an appropriate connection to reach the other machine.
If you need optional GPU components, install and test them before travelling. A working CPU or other supported route may be sufficient for your task, so judge the actual result rather than delaying useful work for an untested setting.
How do you prepare your documents?
Save real local copies, not just shortcuts to cloud files. Open each one before disconnecting to confirm it is available. Keep the original as well as any checked text version you need for the model.
TXT, Markdown, DOCX, and text PDFs are useful for document work. A scanned PDF may need readable text prepared beforehand. The file-processing path extracts source text; it does not fetch missing cloud content for you offline.
Use clear filenames so you can select the right version quickly. For the discovery example, save the approved brief, your current question list, and any reference notes you plan to use.
If you want project search, create the project and index its files while connected. Wait for indexing to finish and ask a known-answer question before departure.
What is a proper disconnected test?
Disconnect the computer from Wi-Fi and Ethernet, then run the exact workflow you want to use. A model responding to “hello” does not establish that your document extraction or audio path is also ready.
For the client-brief example:
- Start a new local chat while disconnected.
- Use + > Attach files to select the saved brief.
- Check the extracted text preview.
- Ask for discovery questions based only on that source.
- Save the useful output through your normal local document workflow.
Restart the app and repeat a short request if you want to check that setup persists. Keep the test small enough that you can diagnose any missing resource before leaving.
What should you do if the test fails?
Separate missing resources from a task that is too large. Reconnect while you still can, complete the needed download, and repeat the disconnected test afterward.
| Failure | What to check |
|---|---|
| Chat cannot start | Local model download and selection |
| A document is unavailable | Whether it is a real local copy |
| Project search is not ready | Indexing resources and completed file indexing |
| Speech processing fails | Local speech model and runtime setup |
| A request waits for a server | Whether a remote model or connected tool is selected |
If a larger task fails, try a shorter source or smaller model. The first successful request is a baseline, not proof that every possible workload will fit the available memory.
How should you plan the work session?
Choose tasks that tolerate being disconnected. Drafting from saved material, outlining, and reviewing known sources are better fits than checking current prices, live availability, or new information from the web.
Keep an explicit list of items to verify later. If a draft needs a current fact, mark it rather than allowing the model to fill the gap from its training data.
Consider battery and workspace limits in your own plan. Local inference uses the computer’s resources, and the app does not guarantee a particular battery duration. Follow the airline’s rules for device use during the journey.
Leave with one verified offline workflow
Download OGAD before the trip, prepare one local model and a small source collection, and test the complete task without a connection.
The useful result is confidence based on your own test: the files open, the model runs, and the output is worth reviewing. When you board, you can start the work rather than troubleshooting the setup.