requirements.txt

# Case Zero - Hugging Face Space (Docker SDK, CPU). Python 3.12.
# Every model is open-weights and self-run. Total runtime params ~1.6B
# (Qwen2.5-1.5B + Supertonic ~0.1B), far under the 32B cap (<=4B Tiny Titan).

# LLM runtime: in-process llama.cpp, compiled from source in the Docker build (PyPI ships
# only an sdist; the prebuilt linux wheels are musl-only and fail on glibc). The Dockerfile
# supplies gcc + cmake and a portable (non-native) build; the compiled layer is cached.
llama-cpp-python==0.3.26

# gradio 6 ships `gradio.Server` (FastAPI-based custom-frontend / "Server mode"), which
# serves the Preact pixel-art SPA + the /api routes from one process. It relaxes the old
# pydantic<2.12 cap.
gradio>=6.16,<7
pydantic>=2.13,<3
pydantic-settings>=2.3,<3
pillow>=11.3,<13
numpy==2.2.6

# TTS: Supertonic (on-device ONNX voices). Pulls onnxruntime.
supertonic==1.3.1

# Used at startup to fetch the open GGUF weights onto the Space.
huggingface_hub==0.36.0