Google launches Gemma 3n, a new AI model for on-device multimodal applications

Google has announced the full release of Gemma 3n, a new open AI model designed to run efficiently on phones, tablets, and laptops. This release marks a substantial advance for on-device AI, as Gemma 3n delivers native support for image, audio, video, and text inputs, while generating text outputs. Previously, such multimodal capabilities were only available in last year’s cloud-based models, but Gemma 3n brings this performance directly to edge devices.

Building on this technical foundation, Gemma 3n integrates with widely used AI tools such as Hugging Face Transformers, llama.cpp, Google AI Edge, Ollama, and MLX, enabling users to fine-tune and deploy models quickly for their specific needs. These features are complemented by two model sizes, E2B and E4B, which require as little as 2 GB or 3 GB of memory, thanks to architectural advances that minimize their operational footprint.

At its core, the model includes several novel components: the MatFormer architecture for improved computational flexibility, Per Layer Embeddings for memory efficiency, and MobileNet-v5 based vision encoders optimized for mobile use. Gemma 3n improves on previous models with expanded multilingual support — 140 languages for text, 35 for multimodal understanding — and offers enhanced math, coding, and reasoning abilities. The models are now available via Google AI Studio, Hugging Face, Kaggle, and other platforms.