I use Ollama which offloads to the CPU automatically IIRC. IME the performance drops dramatically when that happens, and it hogs the CPU making the system unresponsive for other tasks, so I try to avoid it.

I don't believe that's the same thing. That should be the generic offloading that ollama will do to any too big model, while this feature requires MoE models. https://github.com/ollama/ollama/issues/11772 is the feature request for similar on ollama.

One comment in that thread mentions getting almost 30tk/s from gpt-oss-120b on a 3090 with llama.cpp compared to 8tk/s with ollama.

This feature is limited to MoE models, but those seem to be gaining traction with gpt-oss, glm-4.5, and qwen3