Gemma 3n
Mobile-first open model family for on-device text, image and audio work
Gemma 3n is a preview open model family from Google DeepMind, announced at Google I/O 2025, built for phones, tablets and laptops. It uses Per-Layer Embeddings and MatFormer training to run 5B and 8B raw parameter models with memory footprints of 2GB and 3GB, and shares its architecture with the next generation of Gemini Nano.
- Raw parameter counts of 5B and 8B, with 2GB and 3GB dynamic memory footprints
- Per-Layer Embeddings cut RAM use to levels comparable with 2B and 4B models
- MatFormer training nests a 2B submodel inside the 4B model
- About 1.5x faster first response on mobile than Gemma 3 4B, per Google
- 50.1% ChrF on WMT24++ translation, as reported
What it is
Gemma 3n is an early preview of an open model family from Google DeepMind, announced in May 2025 alongside Google I/O 2025. It is designed to run directly on phones, tablets and laptops rather than in a data centre. The raw parameter counts are 5B and 8B, but a technique the company calls Per-Layer Embeddings cuts memory use so the models behave with overheads closer to 2B and 4B models, giving dynamic memory footprints of 2GB and 3GB. It handles text, images, audio and video, and is aimed at real-time, private, offline-capable applications.
How it came about
Gemma 3n follows the Gemma 3 and Gemma 3 QAT releases and shares its underlying architecture with the next generation of Gemini Nano, which Google says will reach features in Google apps later in 2025. The architecture was developed with mobile hardware partners Qualcomm Technologies, MediaTek and Samsung's System LSI business. Google DeepMind describes the release as a preview, with a technical report on its mix'n'match research promised later.
Inside the model
The family uses a mobile-first architecture built around Per-Layer Embeddings (PLE), which the company says greatly reduces RAM use, plus KVC sharing and advanced activation quantisation. MatFormer training nests a 2B active memory footprint submodel inside the 4B one, so a single hosted model can trade quality against latency. A mix'n'match feature lets developers cut further submodels from the 4B model to suit a use case. The models accept interleaved inputs across modalities. Google says the same foundation will reach Android and Chrome.
What it can do
Google reports that Gemma 3n begins responding about 1.5 times faster on mobile than Gemma 3 4B, with better quality and a smaller memory footprint. It does automatic speech recognition and speech-to-text translation, and the company claims improved video understanding and stronger multilingual results, naming Japanese, German, Korean, Spanish and French. One reported figure is 50.1% ChrF on the WMT24++ translation benchmark. A Chatbot Arena Elo chart is cited as placing it well among proprietary and open models.
How to use it
The preview is available in Google AI Studio for browser-based exploration of text input, and through Google AI Edge tools and libraries for local development with text and image capabilities. Audio support is described as coming soon. Google says the architecture will progressively reach Android and Chrome.
Safety and testing
Google states that Gemma 3n went through the same safety evaluations, data governance and fine-tuning alignment with its safety policies as other Gemma models. The company says it applies careful risk assessment to open models and keeps refining its practices.
Why it matters
Google positions Gemma 3n as a step towards making efficient AI widely accessible, and as the open route onto a foundation that will also power the next Gemini Nano across Android and Chrome. Hardware partners named in the collaboration are Qualcomm Technologies, MediaTek and Samsung's System LSI business.
Questions people ask about Gemma 3n
What is Gemma 3n?
It is a preview open model family from Google DeepMind, announced in May 2025, built to run on phones, tablets and laptops. It handles text, images, audio and video, and is meant for fast, private, offline-capable applications on everyday devices.
How large are the models?
The raw parameter counts are 5B and 8B. Per-Layer Embeddings reduce the memory overhead so they run with footprints comparable to 2B and 4B models, giving dynamic memory use of 2GB and 3GB according to Google.
What can it do with audio?
Google says it performs automatic speech recognition and speech-to-text translation at high quality, and accepts interleaved inputs across audio, text and images. The public implementation of the audio capabilities was described as coming soon at announcement.
How do I try it?
At announcement, the preview could be run in Google AI Studio in the browser for text input, or integrated locally with Google AI Edge tools and libraries for text and image work. Google says wider availability on Android and Chrome will follow.
How does it relate to Gemini Nano?
Gemma 3n is the first open model on an architecture that also powers the next generation of Gemini Nano. Google says that version will appear in features across Google apps and its on-device ecosystem later in 2025.
Good reading and sources
There is no English Wikipedia article on Gemma 3n yet, so the facts on this page come from the sources listed above: developers.googleblog.com/en/introducing-gemma-3n/. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.