Gemma: Google unveils open AI models

Includes safety tools as standard

Google has launched Gemma, a family of open AI models designed to empower individuals and businesses to create their own AI software, while also fostering responsible development.

Gemma models, developed by Google DeepMind in collaboration with other teams across Google, are built upon the same cutting-edge technology as the acclaimed Gemini models.

The models are available in two sizes, Gemma 2B and Gemma 7B, with parameters ranging from 2 billion to 7 billion.

Each size boasts pre-trained and instruction-tuned variants, catering to a range of application needs.

One of Gemma's key features is its smaller model sizes, enabling it to run on laptops and computers without sacrificing performance.

Google says Gemma outperforms larger models in "key benchmarks," making it a potential option for developers seeking optimal performance in AI applications.

Google has optimised the new models for seamless integration across popular AI frameworks like JAX, PyTorch and TensorFlow, with native support for Keras 3.0. This ensures maximum flexibility and accessibility for developers, regardless of their preferred toolset.

To ensure the responsible use of AI, Google is also introducing the Responsible Generative AI Toolkit, a suite of tools to guide developers and researchers in creating AI applications that prioritise safety and ethical considerations.

Included in the toolkit are safety classification methodologies, a model debugging tool and best practices for model builders, providing developers with the necessary guidance to navigate ethical challenges in AI development.

Despite that, Gemma still has its share of controversies and inconsistencies - like most gen-AI systems. Today, Google had to pause Gemma's ability to create images of people due to racial inaccuracies.

"We're already working to address recent issues with Gemini's image generation feature," Google said in a post on X. "We're going to pause the image generation of people and will re-release an improved version soon."

Lowering barriers to entry

"We believe in the power of AI to transform industries and drive innovation," Google said.

"With Gemma, we aim to lower the barriers to entry for AI development, empowering developers of all backgrounds to harness the potential of machine learning."

Gemma is not fully open source, and Google has retained some control over the terms of use and ownership.

Nvidia has worked with with Google to ensure the integration of Gemma models with its chips, further enhancing accessibility and performance. The company has also announced plans to make its chatbot software compatible with Gemma.

To encourage widespread adoption and collaboration within the developer community, Google is offering free access to Gemma on platforms such as Kaggle and Colab, along with $300 in credits for first-time Google Cloud users (a standard way cloud providers tempt new customers, aiming to get them entrenched in the ecosystem).

Researchers can also apply for up to $500,000 in free Google Cloud credits, fostering innovation and exploration in AI research and development.

Google also announced a significant update to its AI capabilities within Google Workspace this week.

Duet AI has been rebranded as "Gemini for Google Workspace," granting users access to Google's most advanced AI models. The company says this integration will incorporate Gemini into the daily workflow of millions of Workspace users worldwide - much as Microsoft is hoping to do with Copilot and Office.

Gemini for Workspace is available through two plans.

The Gemini Enterprise add-on replaces the previous Duet AI for Workspace Enterprise offering, providing access to advanced AI capabilities for larger organisations.

Gemini Business, available at a reduced price point, extends the benefits of generative AI to teams and organisations of all sizes, facilitating broader adoption and integration within the Workspace ecosystem.