A tiny AI that lives inside your spruceOS handheld.
0.5 billion parameters. No internet. No cloud. Just you and a little language model vibing on a handheld gaming device.
| Device | Architecture | Status |
|---|---|---|
| Miyoo A30 | armhf (32-bit) | Tested |
| Miyoo Flip | aarch64 (64-bit) | Tested |
| Trimui Brick | aarch64 (64-bit) | Tested |
| Trimui Smart Pro | aarch64 (64-bit) | Tested |
| Trimui Smart Pro S | aarch64 (64-bit) | Tested |
SpruceChat runs Qwen2.5-0.5B entirely on-device using llama.cpp. A persistent server keeps the model loaded in RAM so after the first boot (~60s), each message just... goes. Tokens stream in one by one so you can watch it think.
The AI has the personality of a spruce tree. Patient. Unhurried. Quietly amazed by everything.
Download the latest release — it includes everything (both architectures and the model). Unzip SpruceChat/ to /mnt/SDCARD/App/ on your SD card and launch from the Apps menu.
First boot takes about a minute while the model loads into RAM. After that, you're chatting.
| Button | What it does |
|---|---|
| A | Open keyboard / type selected key |
| B | Backspace / close keyboard / quit |
| X | Space |
| Y / START | Send message |
| L1 | Shift |
| R1 | Delete |
| SELECT | Clear chat history |
| MENU | Quit |
| D-pad | Navigate keyboard / scroll chat |
The llama-server listens on all interfaces (0.0.0.0:8086). If your device is on wifi, you can open http://<device-ip>:8086 in a browser on any device on the same network and chat through llama-server's built-in web UI. The AI runs on the handheld but you get a real keyboard.
On the Miyoo A30 (Cortex-A7, quad-core):
- Model load: ~60s (one time per launch)
- Prompt eval: ~3 tokens/sec
- Generation: ~1-2 tokens/sec
64-bit devices (Flip, Brick, Smart Pro) are faster.
It's not fast. But it streams, so you see each word appear as it thinks. A short response takes 10-30s.
launch.sh detects the device platform and starts the appropriate binary (llama-server for 64-bit, llama-server32 for A30). The Python UI connects over localhost HTTP. The model stays in RAM between messages.
The app ships two sets of binaries and libraries:
llama-server+lib/— aarch64 (Flip, Brick, Smart Pro, Smart Pro S)llama-server32+lib32/— armhf (A30)
Screen resolution and rotation are read from the spruceOS platform config. Input key codes are also read from the platform config, so controls work across all devices.
Builds run on GitHub Actions. Trigger manually from the Actions tab.
The workflow builds both architectures in parallel using Docker:
- A30: steward-fu A30 buildroot toolchain (GCC 13.2.0, glibc 2.23 sysroot)
- Universal: Ubuntu 20.04 aarch64-linux-gnu cross-compiler
Both builds use ccache (persisted between runs) and Docker images cached on GHCR.
To build locally, see Dockerfile.a30 and Dockerfile.universal for the build environments, and build-a30.sh / build-universal.sh for the build scripts.
Originally built by Cassius Oldenburg. Multi-device support by Sundownersport.
MIT