Google Introduces Gemma, a Family of Lightweight and Open AI Models

Google today announced Gemma, a new family of lightweight, open AI models that is built on the same technologies as Gemini.

“We have a long history of contributing innovations to the open community, such as with Transformers, TensorFlow, BERT, T5, JAX, AlphaFold, and AlphaCode,” a new post to the Google for Developers blog notes. “Today, we’re excited to introduce a new generation of open models from Google to assist developers and researchers in building AI responsibly.”

Like Gemini, Gemma was created by Google DeepMind and other teams across the company. And as you might have guessed, its name was inspired by Gemini: In Latin, Gemma means “precious stone.” But unlike Gemini, Gemma is designed to run locally on laptops, PCs, and workstations in addition to Google Cloud, and the Gemma models are optimized for multiple AI platforms, including NVIDIA GPUs and Google Cloud TPUs. Gemma is also licensed for commercial use.

Gemma is available in two model weights, Gemma 2B and Gemma 7B, and each is available in pre-trained and instruction-tuned variants. Google is also providing a responsible generative AI toolkit for Gemma, inference and supervised fine-tuning (SFT) toolchains across all major frameworks (JAX, PyTorch, and TensorFlow through native Keras 3.0), and ready-to-use Colab and Kaggle notebooks. And the models integrate with popular tools Hugging Face, MaxText, and NVIDIA NeMo so you can get up and running quickly.

According to Google, the Gemma models provide “best-in-class performance for their sizes compared to other open models” and outperform “significantly larger models on key benchmarks while adhering to [Google’s] rigorous standards for safe and responsible outputs.” Google plans to expand the Gemma family in the future with new models that target specific use cases.

You can learn more on the Gemma website.