Gemma 3n E2B

Gemma 3n E2B was released by Google on June 26, 2025 as the smaller tier of the Gemma 3n family, designed for on-device multimodal AI. Its MatFormer (Matryoshka Transformer) architecture nests a functional E2B sub-model inside the larger E4B, and per-layer embeddings plus KV cache sharing let it run with as little as 2GB of accelerator memory while handling text, image, audio, and video input.\n\nThe weights are openly downloadable under Google's Gemma Terms of Use, a responsible-use license that allows fine-tuning and commercial deployment but is not an OSI open-source license. It supports 140 languages for text and a 32K-token context, making it a strong fit for privacy-preserving on-device assistants.