Google has officially launched Gemma 3n, the latest addition to its open-source Gemma 3 AI model family, designed specifically for on-device use with as little as 2GB RAM. This breakthrough makes it one of the most lightweight yet capable AI models available, potentially usable on smartphones and other low-resource hardware.
A Multimodal, Mobile-First AI Model
Gemma 3n is a multimodal AI model, meaning it can process images, audio, video, and text as input, while generating only text outputs. The model is multilingual, supporting text in 140 languages and multimodal inputs in 35 languages, making it suitable for a wide range of global applications.
Powered by MatFormer: A Nested Transformer Architecture
Built on Google’s new Matryoshka Transformer (MatFormer) architecture, Gemma 3n uses a nested transformer design—much like Russian nesting dolls—allowing for simultaneous training of different parameter models.
- Gemma 3n E2B has 2 billion effective parameters
- Gemma 3n E4B has 4 billion effective parameters
Though the total size of the models may be 5B and 8B respectively, only the essential parameters are activated during execution, thanks to a technique called Per-Layer Embeddings (PLE).
Customization and Developer Tools
For developers, Google introduced MatFormer Lab, a tool that lets users customize model sizes by tweaking internal parameters—ideal for building task-specific lightweight AI systems.
Developers can access the model through:
Gemma 3n models can also be directly deployed to Cloud Run via AI Studio.
Why Gemma 3n Matters
With increasing demand for AI that works offline or on low-resource devices, Google’s Gemma 3n sets a new standard for efficiency without compromising power. It empowers developers and businesses to create AI applications that are faster, more secure, and more private, all while running locally on edge devices.
