Google has released an early preview version of Gemma 3n, an open source AI model that incorporates optimization techniques for smartphones. There is already a simple way to test it on an Android smartphone, so I tried it out.
Gemma 3n – Google DeepMind
https://deepmind.google/models/gemma/gemma-3n/
Announcing Gemma 3n preview: powerful, efficient, mobile-first AI – Google Developers Blog
https://developers.googleblog.com/en/introducing-gemma-3n/
Gemma 3n is an AI model developed with the intention of running locally on smartphones and tablets, and employs a technology called ‘ Per-Layer Embeddings (PLE) ‘ that can significantly reduce memory usage. Gemma 3n comes in two parameter numbers, 5B and 8B, with the memory usage of the Gemma 3n 5B model being equivalent to a typical 2B model, and the memory usage of the Gemma 3n 8B model being equivalent to a typical 4B model.
Below is a graph comparing the scores in ‘Chatbot Arena,’ which allows humans to evaluate the performance of chat AI. Gemma 3n recorded scores that exceeded ‘ GPT-4.1 nano ‘ and ‘ Llama-4-Maverick-17B-128E-Instruct .’
I heard that you can easily try out Gemma 3n by using the Google Android app ‘
Google AI Edge Gallery ‘, so I actually installed Google AI Edge Gallery on my Pixel 8 Pro and ran Gemma 3n to chat. First, click the link below to open the Google AI Edge Gallery distribution page.
Releases · google-ai-edge/gallery · GitHub
https://github.com/google-ai-edge/gallery/releases
Once you open the distribution page, tap the latest version of ‘ai-edge-gallery.apk’ to download it. This time, I downloaded version 1.0.0.
When the warning appears, tap ‘Continue Download’.
Tap ‘Open’.
Tap ‘Install’.
Once the installation is complete, tap ‘Open.’
The home screen of the Google AI Edge Gallery looks like this. This time, I want to chat with Gemma 3n, so I tap ‘AI Chat’.
When the model selection screen appears, tap ‘Gemma-3n-E2B-it-int4’.
Tap ‘Download & Try’.
When the Hugging Face login screen appears, enter your email address and password and tap ‘Login.’ If you don’t have an account, tap ‘Sign Up’ to create a new one.
Tap “Authorize”.
When you are prompted to agree to the user terms, tap ‘Open user agreement.’
When you see the following screen, scroll down.
Read the terms and license carefully and then tap ‘Agree and send request to access rep.’
Tap “Authorize”.
This will bring up the Access Request screen.
Enter your first and last name and read the terms and conditions.
Check the box to agree to the terms and conditions, then tap ‘Accept.’
When the following screen appears, tap the ‘X’ in the upper left corner.
This will start downloading the model data.
The download took about 2 minutes. After that, just enter text in the input field at the bottom of the screen and Gemma 3n will respond.
Try typing, ‘Can you give me a recipe for something I can make with green onions and carrots?’ and tapping the send button.
It gave me recipes for stir-fry and miso soup. The text generation process runs on the Pixel 8 Pro device, so it can be used offline.
Gemma 3n also supports simultaneous input of images and text, allowing you to ‘ask for image details.’ I want to try it out, so I tapped ‘Ask Image.’
When the model selection screen appears, tap ‘Gemma-3n-E2B-it-int4’.
Tap ‘Try it’.
The chat screen is displayed. Since the model download was completed when I tried the text chat, there is no need for any troublesome consent operations this time. I want to ask about the photo I took with the camera, so I tap ‘+’ at the bottom left of the screen and then tap ‘Take a picture’.
When asked for permission to use the camera, tap ‘Only while using the app.’
Take a picture of the item you want to know more about. This time, I took a picture of the iPhone 13 Pro.
Enter ‘What is this’ and tap the send button.
The answer I got was ‘iPhone 13’. They said they judged it to be ‘iPhone 13’ based on the camera and logo. It was correct to say it was an iPhone, but for iPhones from the iPhone 11 series onwards, there is a way to tell if it is a Pro model if it has three cameras, so I would have liked them to judge whether it was a Pro model.
Google AI Edge Gallery is also developing an iOS version. The memory usage reduction technology used in Gemma 3n will also be used in the next-generation Gemini Nano, scheduled for release in late 2025.
