ParentFull threadnolist_policy·FWIW, I can run Gemma-3-12b-it-qat on my Galaxy Fold 4 with 12Gb ram at around 1.5 tokens / s. I use plain llama.cpp with Termux.View on HN