Google has just launched Gemma 3, its new collection of lightweight, state-of-the-art open-source artificial intelligence (AI) development models based on the same technology powering the Gemini 2.0, and here’s how you can get onboard.
As it happens, Google announced it was introducing Gemma 3, its “most advanced, portable, and responsibly developed open models yet, (…) helping developers create AI applications, wherever people need them,” according to the company’s press release from March 12.
What is Gemma in AI?
Introduced in February 2024, Gemma is the result of work of Google DeepMind and other teams across Google and inspired by Gemini, with which it shares technical and infrastructure components. It refers to Google’s family of pre-trained and instruction-tuned open models that can run on laptops, workstations, and Google Cloud.
The new Gemma 3 comes in four distinct sizes: 1B, 4B, 12B, and 27B parameters. The latter three variants are capable of processing both images and text via a SigLIP encoder and feature support for over 140 languages. In terms of hosting, all the models use the Hub and integration with Hugging Face.
What is the function of Gemma?
Relying on the same research and tech used to create Gemini models, Gemma’s primary role is to generate text tokenword by tokenword based on a user-provided prompt. For instance, in a translation task, Gemma will take a sentence from one language as input and produce output as its equivalent in a different language.
Developers can use the Gemma 3 models for a wide array of text generation and image-understanding tasks and applications, such as answering questions, summarization, and reasoning. They also support function calling and structured outputs to automate workflows more easily.
Is Gemma like ChatGPT?
No. Both Gemini and ChatGPT are closed models, which means that developers cannot modify their code, unlike with the open Gemma. Moreover, neither of them is light enough to run on machines like laptops. It is more like Meta’s Llama, which is an open-source model that can operate on laptops.
How to run Gemma 3?
There are several ways in which you can access Gemma 3. You can try it out directly and setup-free in the browser through Google AI Studio, acquire an API key from Google AI Studio and integrate Gemma 3 into your applications via the Google GenAI SDK, or customize the models through platforms like Hugging Face, Ollama, or Kaggle.
For instance, if you want to run Gemma 3 on your local machine using Ollama, you’ll need to install the framework, use the command-line interface to pull the Gemma 3 variant you need, like “ollama pull gemma3:4b”, start the model locally by executing “ollama run gemma3:4b”, and interact with it directly through your terminal or any local interface from Ollama.
You can also run it on your system or use Google Colab with Hugging Face’s support. This involves installing Python installed along with necessary libraries if on a local system or opening a new notebook and enabling GPU acceleration from the runtime settings if using Google Colab, with detailed instructions provided by Analytics Vidhya.
Elsewhere, Google introduced Gemini Robotics, a new AI model for building robots that draws upon on Gemini 2.0. It is an advanced vision-language-action (VLA) model built on physical actions as a new output modality for the purpose of directly controlling robots.
ReadNext
