Vulkan numbers with the batched Gemma-4 prompt graph (cortiq 0.8.16)

Files carry max_tokens=2048 in their sampling defaults (cortiq 0.8.16 serve uses it); weights unchanged

Metal numbers with the Gemma-4 graph (cortiq 0.8.16)

Model card

humanizer 12B in CMF: q8_2f (default) and q4tp-ffn

initial commit