Google has announced the launch of a new series of open models named Gemma, designed to assist developers and researchers in the responsible creation of artificial intelligence. The Gemma 2b and Gemma 7b models, with 2 and 7 billion parameters respectively, are available in pre-trained and tuned versions. These models are specifically tailored for text and outperform similar models in terms of effectiveness, while being compatible with various hardware platforms for AI, including NVIDIA GPUs and Google Cloud TPUs.
Prior to testing the models with training data, sensitive information such as personal data was removed. Extensive testing was conducted to minimize any undesirable behavior in the models. The models undergo the MMLU test, which measures their knowledge breadth and problem-solving abilities acquired during training.
Unlike previous releases, Gemma is distributed under an open license that allows developers to customize and enhance the models, as well as utilize tools to ensure application safety and compatibility with leading AI platforms and frameworks. Google asserts that Gemma models can be responsibly utilized for commercial purposes and distributed within organizations of any size.
Google offers tools and resources to configure Gemma models for specific applications, with support for various tools and systems, device compatibility, and optimization for Google Cloud services such as Vertex AI and Google Kubernetes Engine. Developers and researchers can access Gemma through Kaggle, utilize Google Colab, and receive $500,000 in Google Cloud credits to accelerate their projects. Google eagerly anticipates the innovative applications and advancements that the developer and research community will create using these new models.