Gemma 3n E4B Instruct
Google DeepMind · 2025-06-26 · 7.8B parameters
Gemma 3n E4B Instruct is the larger of Google's two Gemma 3n instruction-tuned on-device models, fully released on June 26, 2025. Its MatFormer architecture nests an E2B submodel, while Per-Layer Embeddings let only approximately 4B core transformer parameters reside in accelerator memory even though the checkpoint has about 8B parameters; Google reports operation with as little as 3 GB of memory. It accepts text, image, audio, and video and outputs text, with support for 140 languages in text and multimodal understanding in 35 languages, targeting resource-constrained phones and edge devices.