Brattoi: a tiny AI potato that runs on your muOS handheld

Applications Release
1

Brattoi

Brattoi is an app that runs a small language model on the handheld itself, inside Mustard Terminal. No cloud, no account. It is a demo of what the hardware and muOS can do.

What it does:

  • Ask muOS: ready-made help questions, answered from a bundled copy of the muOS docs and forum FAQ as of October 2026. Sources shown.
  • Ask your own question: type one with the on-screen keyboard.
  • Just for fun: silly prompts.
  • Free chat: talk to it.
  • Models: download, switch or delete models.

It is a very small model doing its best. It makes mistakes and says silly things. For real help, go to https://muos.dev.

Download

Install

  1. Copy Brattoi_v0.12.muxapp to ARCHIVE/ on the SD card.
  2. Install it with Archive Manager. Brattoi shows up in Applications.
  3. Launch it. The first launch shows the controls.
  4. Download a model from Models. Needs WiFi.

No WiFi on your device (the Pixel 2 has none built in)? Download the model on a PC from the link in models.json and copy the .gguf file to MUOS/application/Brattoi/.brattoi/ on the SD card.

On my GKD Pixel 2 I used this to provide WiFi and BT: https://www.aliexpress.com/item/1005012652537014.html

Which model

Get Granite 4.0 350M and Qwen2.5 0.5B and pick Auto. Granite answers muOS questions, Qwen does the chats. About 5 to 7 words a second on the Pixel 2. A help answer takes 30 to 60 seconds.

Models

Model Size Good at Direct download
Granite 4.0 350M 228 MB Help answers (recommended) https://huggingface.co/ibm-granite/granite-4.0-350m-GGUF/resolve/main/granite-4.0-350m-Q4_0.gguf
Qwen2.5 0.5B 353 MB Chat https://huggingface.co/bartowski/Qwen2.5-0.5B-Instruct-GGUF/resolve/main/Qwen2.5-0.5B-Instruct-Q4_0.gguf
LFM2 350M 219 MB Some sense https://huggingface.co/LiquidAI/LFM2-350M-GGUF/resolve/main/LFM2-350M-Q4_0.gguf
SmolLM2 360M 230 MB Some sense https://huggingface.co/bartowski/SmolLM2-360M-Instruct-GGUF/resolve/main/SmolLM2-360M-Instruct-Q4_0.gguf
Gemma 3 270M 242 MB Not accurate https://huggingface.co/unsloth/gemma-3-270m-it-GGUF/resolve/main/gemma-3-270m-it-Q4_0.gguf
SmolLM2 135M 92 MB Not accurate https://huggingface.co/bartowski/SmolLM2-135M-Instruct-GGUF/resolve/main/SmolLM2-135M-Instruct-Q4_0.gguf

Devices

Tested on the GKD Pixel 2 on muOS 2606.0 Andromeda. Should run on other aarch64 muOS devices, untested. Reports welcome, here or on GitHub.

Credits

Logo: @PotatoJesus “Brotato’s” avatar, used with their OK. The name came from SmolLM2 calling itself “Brattoi”. Runs on llama.cpp. Licences in the repo.

3 Likes