Google has launched Gemma, a new family of lightweight open models that includes Gemma 2B and Gemma 7B. These models were inspired by Google's Gemini models and are designed for both commercial and research use. While Google describes Gemma models as 'state-of-the-art,' there is a lack of detailed information on how they perform compared to similar models from competitors like Meta and Mistral.

Gemma models follow the same architecture as Google's Gemini and earlier PaLM models, using a dense decoder-only approach. The benchmarks for Gemma models will be released on Hugging Face's leaderboard. To assist developers in getting started with Gemma, Google provides ready-to-use Colab and Kaggle notebooks, as well as integrations with Hugging Face, MaxText, and Nvidia's NeMo. Once pre-trained and fine-tuned, these models can be deployed in various environments.

Why does it matter?

Although Google emphasises that Gemma models are open models, they are not open-source. Google's Janine Banks explained that the models are referred to as 'open models' instead of 'open source' because the specific terms of use, including redistribution and ownership of variants, vary. Nonetheless, developers can still use Gemma models for inferencing and fine-tuning purposes.

In addition to the Gemma models, Google is also introducing a responsible generative AI toolkit, which provides guidance and tools for creating safer AI applications using Gemma. Furthermore, a debugging tool will be made available to support developers in their AI application development with Gemma models.