Google introduced Gemma, a state-of-the-art open-source model family

Gemma comes in 7B and 2B model weights, and each is released with a pre-trained and an instruction-tuned variant. The models' release includes several developer tools, including inference and supervised fine-tuning toolchains, Colab and Kaggle notebooks, integration with popular tools such as Hugging Face, and most importantly, Google's new Responsible Generative AI Toolkit. Unlike its proprietary solutions, Gemma's license allows responsible commercial use and distribution in organizations of all sizes.

The Gemma models share some technical and infrastructure features with Gemini, which enable the former to deliver state-of-the-art performance surpassing similarly-sized models such as Mistral and Llama-2. Gemma's technical report details insights such as performance, dataset composition, and modeling methods. Gemma models were trained with data filtered of any sensitive or personally identifying information. Additionally, the models have been extensively fine-tuned and aligned via reinforcement learning from human feedback to guarantee their safety and reliability. Other safety evaluations include manual red-teaming, automated adversarial testing, and dangerous activity assessments.

The Gemma models can be used starting February 21 via Kaggle, a free tier for Colab notebooks and $300 in Google Cloud credits for first-time users. Research teams and individuals can also apply for additional $5K credit grants (up to a collective $500,000 between all selected applicants).