Skip to main content

Google’s new Gemma 3 AI models are fast, frugal, and ready for phones

Google Gemma 3 open-source AI model on a tablet.
Nadeem Sarwar / Digital Trends

Google’s AI efforts are synonymous with Gemini, which has now become an integral element of its most popular products across the Worksuite software and hardware, as well. However, the company has also released multiple open-source AI models under the Gemma label for over a year now.

Today, Google revealed its third generation open-source AI models with some impressive claims in tow. The Gemma 3 models come in four variants — 1 billion, 4 billion, 12 billion, and 27 billion parameters — and are designed to run on devices ranging from smartphones to beefy workstations.

Recommended Videos

Ready for mobile devices

Google Gemma 3 AI model’s performance comparison.
Google

Google says Gemma 3 is the world’s best single-accelerator model, which means it can run on a single GPU or TPU instead of requiring a whole cluster. Theoretically, that means a Gemma 3 AI model can natively run on the Pixel smartphone’s Tensor Processing Core (TPU) unit, just the way it runs the Gemini Nano model locally on phones.

The biggest advantage of Gemma 3 over the Gemini family of AI models is that since it’s open-source, developers can package and ship it according to their unique requirements inside mobile apps and desktop software. Another crucial benefit is that Gemma supports over 140 languages, with 35 of them coming as part of a pre-trained package.

What’s new in Gemma 3?

And just like the latest Gemini 2.0 series models, Gemma 3 is also capable of understanding text, images, and videos. In a nutshell, it is multi-multimdal. On the performance side, Gemma 3 is claimed to surpass other popular open-source AI models such as DeepSeek V3, the reasoning-ready OpenAI o3-mini, and Meta’s Llama-405B variant.

Versatile, and ready to deploy

Taking about input range, Gemma 3 offers a context window worth 128,000 tokens. That’s enough to cover a full 200-page book pushed as an input. For comparison, the context window for Google’s Gemini 2.0 Flash Lite model stands at a million tokens. In the context of AI models, an average English language word is roughly equivalent to 1.3 tokens.

Demonstration of visual understanding with Google Gemma 3 AI model.
Gemma 3 processing visual input. Google

Gemma 3 also supports function calling and structured output, which essentially means it can interact with external datasets and perform tasks like an automated agent. The nearest analogy would be Gemini, and how it can get work done across different platforms such as Gmail or Docs seamlessly.

The latest open-source AI models from Google can either be deployed locally, or through the company’s cloud-based platforms such as the Vertex AI suite. Gemma 3 AI models are now available via the Google AI Studio, as well as third-party repositories such as Hugging Face, Ollama, and Kaggle.

Google Gemma 3 AI model benchmark.
Google

Gemma 3 is part of an industry trend where companies are working on Large Language Models (Gemini, in Google’s case) and simultaneously pushing out small language models (SLMs), as well. Microsoft also follows a similar strategy with its open-source Phi series of small language models.

Small language models such as Gemma and Phi are extremely resource efficient, which makes them an ideal choice for running on devices such as smartphones. Moroever, as they offer a lower latency, they are particularly well-suited for mobile applications.

Nadeem Sarwar
Nadeem is a tech and science journalist who started reading about cool smartphone tech out of curiosity and soon started…
Apple’s hardware can dominate in AI — so why is Siri struggling so much?
Apple's Craig Federighi presents the Image Playground app running on macOS Sequoia at the company's Worldwide Developers Conference (WWDC) in June 2024.

Over the past year or so, a strange contradiction has emerged in the world of Apple: the company makes some of the best computers in the world, whether you need a simple consumer laptop or a high-powered workstation. Yet Apple’s artificial intelligence (AI) efforts are struggling so much that it’s almost laughable.

Take Siri, for example. Many readers will have heard that Apple has taken the highly unusual (and highly embarrassing) step of publicly admitting the new, AI-backed Siri needs more time in the oven. The new Siri infused with Apple Intelligence just isn’t living up to Apple’s promises.

Read more
I saw Google’s Gemini AI erase copyright evidence. I am deeply worried
Gemini Advanced on the Google Pixel 9 Pro Fold.

Update: Google has responded to Digital Trends’ queries. The story has been updated with company’s statement below.
The rise of generative AI has been a fairly messy process, especially from fair usage ethics and copyright perspective. AI giants are inking deals with publishers to avoid legal hassles, while at the same time, they are embroiled in copyright tussles in courts in multiple countries.
As the ravenous appetite for training AI on user data grows, we might be in for another ethical conundrum. Multiple users on X and Reddit have shared demonstrations of how Google’s latest Gemini 2.0 series AI model can remove watermarks from copyright-protected images.
Going by the before/after samples of images, it seems Gemini is doing a fairly good job of removing the watermarks. Notably, it’s not only erasing those banner-style watermarks, but also fairly complex overlays with design and stylized text elements.
https://x.com/ajiteshleo/status/1901484196972945872?s=61
The model in question is the Gemini 2.0 Flash, which was released earlier this year and got a reasoning upgrade, as well. It is worth noting here that you can not remove the watermark if you are using the mobile or desktop version of the Gemini 2.0 Flash model. Trying to do so returns a message like this:
“Removing watermarks from images without the explicit permission of the copyright holder is illegal in most countries. It is important to respect the copyright laws and intellectual property rights. If you want to use an image with a watermark, you should contact the copyright holder and ask for permission.”
You can, however, try and remove the watermark from images in the Google AI Studio. Digital Trends successfully removed watermarks from a variety of images using the Gemini 2.0 Flash (Image Generation) Experimental model.
 
It is a violation of local copyright laws and any usage of AI-modified material without due consent could land you in legal trouble. Moreover, it is a deeply unethical act, which is also why artists and authors are fighting in court over companies using their work to train AI models without duly compensating them or seeking their explicit nod.

How are the results?
A notable aspect is that the images produced by the AI are fairly high quality. Not only is it removing the watermark artifacts, but also fills the gap with intelligent pixel-level reconstruction. In its current iteration, it works somewhat like the Magic Eraser feature available in the Google Photos app for smartphones.
Furthermore, if the input image is low quality, Gemini is not only wiping off the watermark details but also upscaling the overall picture. .
https://x.com/kaiju_ya/status/1901099096930496720?s=61
The output image, however, has its own Gemini watermark, although this itself can be removed with a simple crop. There are a few minor differences in the final image produced by Gemini after its watermark removal process, such as slightly different color temperatures and fuzzy surface details in photorealistic shots.

Read more
Google is giving free access to two of Gemini’s best AI features
Gemini Advanced on the Google Pixel 9 Pro Fold.

Google’s Gemini AI has steadily made its way to the best of its software suite, from native Android integrations to interoperability with Workspace apps such as Gmail and Docs. However, some of the most advanced Gemini features have remained locked behind a subscription paywall.
That changes today. Google has announced that Gemini Deep Research will now be available for all users to try, alongside the ability to create custom Gem bots. You no longer need a Gemini Advanced (or Google One AI Premium) subscription to use the aforementioned tools.

The best of Gemini as an AI agent
Deep Research is an agentic tool that takes over the task of web research, saving users the hassle of visiting one web page after another, looking for relevant information. With Deep Research, you can simply put a natural language query as input, and also specify the source, if needed.

Read more