You set a session length at launch (one hour by default), but you're not locked in, you can extend it later from the running service without interrupting it.

Yes, via Customize at the bottom of the package panel. It hands off into the Custom Models flow, carrying the model and the package's launch parameters over as a starting point, and keeps the package's resource minimums as a floor on whatever you size yourself.

The second model replaces the first, and any launch parameters you'd already set for the first one are discarded, not carried over. Selection is single-model, so switching means reconfiguring from scratch.

vLLM and llama.cpp. vLLM needs a CUDA GPU and is the high-throughput option. llama.cpp runs on CPU or GPU and serves quantized (GGUF) models, which is what lets you run something on a node with no free GPU.

Yes, supply your Hugging Face access token at launch. It's held in your browser tab only, encrypted before it reaches the node, and never written into a URL.

Yes. Both vLLM and llama.cpp expose an OpenAI-compatible API, so anything that calls POST /v1/chat/completions works against your service unchanged.