This is not true.

No inference engine does all of:

- Model switching

- Unload after idle

- Dynamic layer offload to CPU to avoid OOM

this can be added to llama.cpp with llama.swap currently so even without Ollama you are not far off