Google introduces a new lightweight open source AI model called Gemma
Google has released an open-source AI model called Gemma, which they say is built using the same research and technology used to develop their Gemini models. The company notes that Gemma is its contribution to the open source community and is designed to help developers “build AI responsibly.” In this regard, Google has also introduced the Toolkit for Responsible Generative AI with Gemma. It contains a debugging tool as well as a guide with best practices for AI development based on Google’s experience.
The company made Gemma available in two sizes – Gemma 2B and Gemma 7B. Both versions have options with pre-trained and instructionally customized models and are light enough to run directly on developers’ laptops or desktops. Google claims that Gemma outperforms significantly larger models in key benchmarks, and that both model sizes outperform other open source models.
In addition, Gemma’s models have been trained to be safe. Google used automated techniques to remove personal information from the data used to train the models and applied a reinforcement learning technique based on human feedback to ensure that Gemma’s instructionally-tuned variants behaved responsibly. Companies and independent developers can use Gemma to build AI-based applications, especially if none of the existing open source models are powerful enough for their purposes.
Google intends to introduce even more Gemma variants in the future for an even more diverse range of applications. Those who want to start working with the models right now can access them through the Kaggle platform, the company’s Colab laptops, or through Google Cloud.
