Introducing Gemma 3n: The developer guide

Simon Willison reports on Google's release of Gemma 3n, an open weights model that natively accepts image, audio, video and text inputs and produces text outputs. The E2B and E4B sizes have raw parameter counts of 5B and 8B but, according to the article, run with memory footprints comparable to 2B and 4B models, needing as little as 2GB and 3GB. Willison tested the 4B model on a Mac laptop: the Ollama version drew an SVG pelican but did not support image or audio input, while the mlx-vlm version transcribed a WAV file correctly and described an image, though it misidentified the picture as a chemical diagram.