I’m running Proxmox and hate it. I still recommend it for what you are trying to do. I think it would work quite nicely. Three of my four nodes have llama.cpp VMs hosting OpenAI-compatible LLM endpoints (llama-server) and I run Claude Code against that using a simple translation proxy.

Proxmox is very opinionated on certain aspects and I much prefer bare metal k8s for my needs.