Thanks for this project. Prioritizing MoE models and adding an intelligent NVMe cache could improve efficiency, especially on the M4 Max where bandwidth makes usage more realistic.