New Delhi: Google has officially rolled out the full version of its mobile-first AI model, Gemma 3n, after offering a preview last month. This latest release aims to bring advanced AI tools directly to smartphones and edge devices, with features that were earlier possible only on large cloud-based models.
Built with developers in mind, Gemma 3n focuses on running smoothly on-device, making it ideal for apps that need speed, privacy, and offline support. The model is supported across popular platforms like Hugging Face, llama.cpp, Ollama, and MLX, helping devs fine-tune and deploy AI in lightweight environments.
Our open source Gemma models are the most powerful single GPU/TPU models out there! Our latest model Gemma 3n has amazing performance, multimodal understanding, & can run with as little as 2GB of memory – perfect for edge devices – enjoy building at https://t.co/Navvgimwjr ! https://t.co/vfeOlnezsn
— Demis Hassabis (@demishassabis) June 26, 2025
What makes Gemma 3n different?
At the heart of Gemma 3n is a clever architecture called MatFormer, short for Matryoshka Transformer. Think of it like those Russian nesting dolls. One model contains smaller versions inside, allowing developers to choose sizes based on their device needs. The model comes in two sizes, E2B and E4B, with memory usage close to 2GB and 3GB, making it manageable on smartphones and embedded systems.
Gemma 3n can process not just text, but also images, audio, and video. The team says it supports text in 140 languages and can understand visual and audio content in 35. The larger E4B model even crossed a 1300 score on the LMArena benchmark, a first for any sub-10B parameter model.
Features packed for edge devices
The new model also includes features like:
- Per-Layer Embeddings (PLE): This helps reduce memory use while keeping performance high.
- KV Cache Sharing: A big upgrade for handling longer text, audio, or video inputs faster.
- Advanced audio tools: Using Google’s Universal Speech Model, it now supports speech-to-text and language translation directly on the device.
- MobileNet-V5 vision encoder: With this, the model can understand video frames and images with high accuracy, processing up to 60 frames per second on devices like Google Pixel.
More tools, better control
One standout update is Mix-n-Match, a method where developers can fine-tune the size of the model depending on their hardware. Google has also released MatFormer Lab, a tool to help tweak model parameters based on benchmark results.
The company says Gemma 3n was built with community help. It’s backed by developers and contributors from groups like NVIDIA, AMD, Docker, RedHat, and Hugging Face.
To encourage adoption, Google also launched the Gemma 3n Impact Challenge with ₹1.3 crore ($150,000) in prize money, inviting devs to build real-world projects using Gemma’s edge AI strengths and submit a compelling video demo.
The model is now available through Google AI Studio, Hugging Face, and other platforms.