Can you describe what is this slightly different approach and why it should work on all models?

Nice! Your stuff ran LLMs extremely well on < $500 boxes (24-32GB ram) with iGPUS before this update.

I’m eager to try it out, especially if 16GB is viable now.