FROM ollama/ollama:latest

# HF Spaces (free CPU) expects port 7860
ENV OLLAMA_HOST=0.0.0.0:7860
ENV OLLAMA_MAX_LOADED_MODELS=1

EXPOSE 7860

# Override default entrypoint so we can run normal commands
ENTRYPOINT []

# Start Ollama server + pull Qwen3 8B (quantized for CPU) + keep alive
CMD ollama serve & sleep 15 && ollama pull qwen3:8b-q4_K_M && echo "✅ Qwen3 8B ready!" && tail -f /dev/null