Gemma 3n is Google's new open model, optimized for on-device multimodal AI. Its novel MatFormer architecture enables powerful yet efficient models (like the 2B/4B variants) that can run locally on phones and laptops. Supports image, audio & video.
Google is back with a major release for their open model family: Gemma 3n. It’s a big step forward for powerful, on-device multimodal AI.
It has the new MatFormer architecture. It's like a Matryoshka doll, a single, larger model contains smaller, fully-functional models inside. This gives developers incredible flexibility. You can deploy a tiny 2B effective parameter model for speed, a more powerful 4B version, or even create custom sizes in between.
And it’s built to be very efficient. Techniques like Per-Layer Embeddings mean only a small part of the model needs to live in VRAM, making it truly viable for phones and laptops. It’s also fully multimodal, handling image, audio, and video inputs with strong performance.
