See this question at least a handful of times when browsing new and in the comments, llamacpp has one of the cleaner model hotswap apis now that just works with openwebui and hermes. Bonus: the 2nd model gemma went derp as i was recording this, but the time spent swapping has gotten stupid fast... I remember starting a load and talking a walk while pytorch did its thing just a few months back podman run -d \ --name llama-qwen36-router \ --device nvidia.com/gpu=all \ -v /data/models:/root/.c...