Procedure Install a Local Model

Procedure 7: Install a Local Model

This procedure puts a chat model on the phone, so chat works offline and costs nothing per message.

Before you start: connect to Wi-Fi and plug in the charger. The suggested model is 2.4 GB. A phone with 8 GB of memory or more works best.

Cost: free. The model comes from Hugging Face, and no key is needed.

Steps

1. On a new install, the empty Chat screen asks Where should Chat run? Tap On this phone and skip to step 4.
2. Otherwise, tap the gear to open Settings, and keep Chat selected.
3. Under Chat runs, tap On this phone, then Manage models.
4. On the first card, Qwen3 4B Instruct, abliterated, tap Download (2.4 GB). The bar shows the progress, and the screen stays awake. On Wi-Fi the download takes about five minutes.
5. Wait for Checking the file. The app compares the file with the checksum Hugging Face publishes, which takes about ten seconds. The card then shows In use.
6. Go back to the main screen. The setup card names the model and says Loads with your first message.
7. Write a message and tap send. The first message loads the model into memory, which takes a few seconds. Later replies start in about a second.

Tips

• Pause stops a download and keeps the part already on the phone. Resume continues from there.
• The line under each reply shows how fast the phone wrote it. The suggested model writes about 7 to 11 tokens per second on a Pixel 10 Pro XL.
• Settings, under Chat, also offers the context size. A larger context remembers more of a long conversation and uses more memory.
• To add a model that is not in the list, open its Files page on Hugging Face, copy the link to one .gguf file, and paste it under Add a model from Hugging Face.
• Delete on a card frees its space. The model can be downloaded again later.