animica-serve

Serve Animica AICF inference jobs from anything with a CPU — phones under Termux included. Torch-free and dependency-free: llama.cpp does the inference, this package does the earning. Each job you win credits its full estimated cost in ANM to your wallet address on the network's worker ledger.

The zero-install browser version of this worker lives at https://pool.animica.org/serve — this package is the terminal lane: it survives reboots in a tmux/sv session, doesn't need the screen on, and uses llama.cpp's native ARM speed.

Termux (Android)

pkg install python llama-cpp        # llama.cpp built for your phone's CPU
pip install animica-serve
animica-serve --address anim1yourwallet

The default model (Qwen 2.5 1.5B Instruct, Q4_K_M, ~1.1 GB) downloads once and is cached; use --model qwen2.5-0.5b on older/low-RAM phones. Add --charge-only to pause while unplugged (needs the Termux:API app + pkg install termux-api). Keep Termux alive in the background with termux-wake-lock.

Anywhere else

Any Linux/macOS box works the same way. If you already run a llama-server or Ollama, skip the model download and point at it:

animica-serve --address anim1… --openai-url http://127.0.0.1:11434/v1 --openai-model qwen2.5:1.5b

No llama-server binary? pip install 'animica-serve[python-backend]' uses llama-cpp-python in-process.

How earning works

Your worker registers your address (no keys ever leave your device), claims jobs from the shared AICF queue on rpc.animica.org, answers locally, and submits. Jobs are raced across several workers — the first good answer wins and is credited (animica-serve earnings --address anim1… shows your ledger). Losing races to faster desktop GPUs is normal; you win whenever you are the fastest or only worker awake.