Omar Sanseviero, a member of Google DeepMind's technical team, has introduced an experiment that brings AI inference to office laser printers. The experiment requires no firmware modifications or additional external compute boards. Using Google's Gemini 4 Carbon model, a small Transformer inference engine and a quantized model are compiled into a Post print job, and the printer's existing Post interpreter executes the model computation, with all computation performed internally on the device by the printer's CPU. The experiment runs the approximately 270-million-parameter Gemma 3 270M model, adapting to limited memory through dynamic layer loading, with peak runtime memory usage of approximately 14.1MB. Model inference involves 18 Transformer network layers and takes about 60 seconds to generate a single token; the prefill stage takes about 184 seconds to process the input prompt, and decoding a single generated token takes 0.1 seconds.
