Google has unveiled Gemma, a pioneering series of open models aimed at empowering developers and researchers to engage in AI development with a conscientious approach. This new suite of models is a collaborative effort by Google DeepMind and various Google teams, drawing inspiration from the Gemini models, emphasizing the importance of responsible AI practices from the ground up.
Gemma models are globally available and come in two variants: Gemma 2B and Gemma 7B, both offering pre-trained and instruction-tuned configurations. To bolster developer creativity and encourage collaborative efforts, Google is also rolling out the Responsible Generative AI Toolkit. This toolkit is a compilation of guidance and crucial tools tailored for crafting safer AI applications utilizing Gemma.
The Gemma models are engineered for seamless integration with leading frameworks, including JAX, PyTorch, and TensorFlow, leveraging the native capabilities of Keras 3.0. Developers and researchers can jumpstart their AI projects with ready-to-use Colab and Kaggle notebooks. Furthermore, Gemma models are designed for compatibility with a wide array of tools such as Hugging Face, MaxText, NVIDIA NeMo, and TensorRT-LLM, ensuring a broad spectrum of deployment possibilities from personal computing devices to cloud-based platforms like Google Cloud, Vertex AI, and Google Kubernetes Engine (GKE). The models are fine-tuned for optimal performance across various AI hardware, including NVIDIA GPUs and Google Cloud TPUs.
When it comes to performance, Gemma models set a new standard for open models of their size, achieving remarkable results while strictly adhering to Google’s stringent guidelines for safe and responsible outputs. The models have undergone comprehensive evaluations to ensure their reliability, including manual red-teaming, automated adversarial testing, and in-depth analysis of their capabilities for potentially hazardous activities.
In the development of Gemma models, Google’s commitment to its AI Principles is evident. Efforts have been made to exclude personal information and sensitive data from the training datasets. Moreover, the models have been fine-tuned with reinforcement learning from human feedback to align them more closely with responsible behaviors. Alongside the models, the Responsible Generative AI Toolkit has been made available, encompassing safety classification, model debugging tools, and guidance grounded in best practices for responsible AI development.
Gemma models offer the flexibility to be fine-tuned on user-specific data and support a wide range of frameworks, tools, and hardware platforms. They can be deployed across various devices including laptops, desktops, IoT devices, mobile devices, and cloud platforms. Notably, these models are optimized for Google Cloud, with integrated support for Vertex AI and customization options to meet different project needs.
Google is democratizing access to Gemma by providing free access through platforms like Kaggle and Colab, in addition to offering credits for first-time Google Cloud users. Researchers seeking to accelerate their projects can also apply for Google Cloud credits. This initiative underscores Google’s commitment to making responsible and accessible AI capabilities available to a broader audience.