flask
gunicorn
jinja2
# Qwen3.8 is a hybrid Gated-DeltaNet architecture: it needs a llama.cpp with the
# qwen3_5/qwen3_8 graph. Use the newest llama-cpp-python, and rebuild against a
# newer llama.cpp (or bump this pin) if the model fails to load.
llama-cpp-python>=0.3.35
# Required: Llama.from_pretrained() downloads the quant through the Hub client.
huggingface-hub
