Off Grid 0.0.116-beta.1 (beta)

Features

• The Models screen now opens with Fast, Balanced and Best picks for your phone. One tap downloads the model and switches to it.
• The Image and Speech tabs also open with Fast, Balanced and Best picks for your phone. On the Image tab, these picks replace the RAM banner.
• Tap "Check for new models" to get newly added models without an app update. A bottom sheet names each new model, and the new models show in the list right away.
• When a chat model finishes downloading, you can tap Chat now. Image models tell you when they are ready.
• Each model's detail view offers its sizes in plain words: Smaller, Recommended and Best. Every file is still under All files.
• When you turn on MTP, the app downloads the model's MTP head for you. You see the progress under the switch, and you can retry with one tap if it fails.
• Your place in a chat is kept when you switch chats or leave the app. When you come back to one of your three most recent chats, it picks up faster.
• "Use fastest settings" tests each backend your phone can run and keeps the fastest. The test takes 30 seconds or less, and you can cancel it.
• Threads and batch size can go back to Auto with one tap.
• Beta builds tell you when a newer engine is available than the one in your build.

Fixes

• If you switch to a model that does not fit in memory, your current model stays loaded. Before, you could lose it.
• When you pick a different LiteRT model, the app now loads that model and not the one already loaded.
• If you cancel the Insufficient Memory warning, your chat stays on the model that is still loaded.
• On an expired Pro install, Back on the Pro screen now takes you home.
• Alerts from a model's detail view now show once and not twice.
• Downloads check free space at the moment they start, including downloads that waited in the queue. If space is low, the app tells you how much it needs and how much is free.
• If recommended model details fail to load, the model cards still show their catalog information.
• Speech model cards keep their place in the list.

Improvements

• Search shows curated models first, then Hugging Face. You get one result per model, with its sizes grouped.
• Search hides models that are too big for your phone or that this version of the app cannot run. Chat search no longer shows image generators.
• Model cards are shorter and show only what helps you choose. For a mixture-of-experts model, the card shows the active size. Cards also tell you when a newer model replaces one, and when a model has a research-only licence.
• Files that are too big for your phone are hidden. Files that only just fit are marked Tight.
• The chat picks come from model makers you know. For example, a 12 GB phone gets Qwen 3.5 0.8B, 2B and 4B.
• Curated models show up under their maker's filter, and Liquid AI now has its own filter.
• Setup never offers the same chat model in two plans. On phones without room for a bigger model, setup shows fewer plans.
• On Android, CPU and NPU users get the Q4_0 version of a model when there is one, because it runs faster there.
• Thread count and batch size are set automatically for your phone, so models load and run faster.
• A model downloaded from the curated list keeps the same name everywhere in the app.
• Screen readers now say which model each download button downloads.
• On Android, LiteRT-LM is updated to 0.18.1.
• Each beta now ships the newest on-device engines.

Build: 1792641997