I would recommend adding the MiniCPM5-1B model. Surprisingly coherent for a 1B model and performs well. On my pixel 9 I get 33 tok/s on CPU. On GPU I get about 26 tok/s but prefill jumps to nearly 500.