So this is a notebook with good enough TPU capabilities to run Gemini partially (like in a MoE), a small model that knows when to delegate to the main model?