dedicated AI gadgets you actually use

[rasterbench]

A couple questions before I guess too hard. What model, what runtime, and roughly what context size?

A couple things I'd check first. Before rebuilding anything, I'd test one clean baseline with Gemma and compare from there.