I tried

    curl -LsSf https://llama.app/install.sh | sh

and then

    llama serve -hf unsloth/Qwen3-4B-GGUF:Q4_0

Then I get:

    W load: control-looking token: 128247 '</s>' was not control-type; this is probably a bug in the model. its type will be overridden
    Terminated

And the web interface says

    Server unavailable

Maybe it gets killed by the OS because it uses too much RAM?

When I try

    llama serve -hf unsloth/Llama-3.2-1B-Instruct-GGUF:Q4_K_M

It seems to work. Nice.