Google releases Gemma lightweight AI open models

Google has introduced Gemma, a new family of open-source lightweight AI models that allow developers and researchers to build large language models (LLM) with only a laptop or desktop.

What makes Gemma most unique is its ability to run across multiple types of devices, including “laptop, desktop, IoT, mobile and cloud, enabling broadly accessible AI capabilities.”

Gemma is said to be born under Google’s new era of Gemini, its latest and most advanced LLM to date and rival ChatGPT competitor, also developed by Google DeepMind using the same research and technology.

A collaboration between Google AI and NVIDIA, Gemma is optimized for NVIDIA GPUs as well as for Google Cloud. The models are offered in two sizes, 2B and 7B, to “achieve best-in-class performance,” Google said.

According to NVIDIA. the state-of-the-art lightweight 2 billion and 7 billion-parameter open models can be run anywhere, "reducing costs and speeding innovative work for domain-specific use cases."

Included with Wednesday’s release is a new Responsible Generative AI Toolkit to provide developers the resources to apply best practices for responsible and safe use of open models like Gemma.

For fine-tuning, Gemma supports a wide variety of multi-framework tools such as “Keras 3.0, native PyTorch, JAX, and Hugging Face Transformers.” It is also equipped with a model debugging tool to help developers “investigate Gemma's behavior and address potential issues,” Google said.

AI toolkit for responsible model creation

The Responsible Generative AI Toolkit is designed to provide the user guidance “on setting safety policies, safety tuning, safety classifiers and model evaluation,” Google states.

It even comes with its own customizable Learning Interpretability Tool (LIT) for investigating Gemma's behavior and addressing potential issues.

The toolkit highlights the following areas of concern when using Gemma to build an AI model.

  • Assess risks and set safety policies
  • Tune models for safety
  • Crete input and output safeguards
  • Evaluate model and system for safety
  • Build transparency artifacts
  • Analyze model behavior

Built for the open developer and research community powering AI innovation, Google says Gemma is now available worldwide through “free access in Kaggle, a free tier for Colab notebooks.”

Google is also offering free credits to from $300 to $500,000 to first time Google Cloud users and researchers looking to accelerator their projects.