Today, Google DeepMind announced the launch of Gemma, its new open-source AI models available in 2B and 7B parameters. These models leverage the same research and technology that powered the recently unveiled Gemini models.
Gemma will come in pre-trained and instruction-tuned variants, accompanied by a permissive commercial license and a Responsible Generative AI toolkit. Additionally, Google DeepMind is providing toolchains for inference and supervised fine-tuning (SFT) compatible with major frameworks: JAX, PyTorch, and TensorFlow via native Keras 3.0. Developers can access ready-to-use Colab and Kaggle notebooks, and Gemma is compatible with Hugging Face, MaxText, and NVIDIA NeMo. The pre-trained and instruction-tuned models can be run on laptops, workstations, or Google Cloud, with deployment options on Vertex AI and Google Kubernetes Engine.
NVIDIA has also collaborated with Google to enhance optimizations across all NVIDIA AI platforms, including local RTX AI PCs, to boost Gemma’s performance.
Jeanine Banks, Google’s Vice President and General Manager of Developer X, noted that the Gemma models represent a continuation of Google’s commitment to open-source technology for AI development, building on tools like TensorFlow and JAX, and models such as PaLM2 and AlphaFold. She highlighted insights gained during the Gemini model development, revealing that developers often utilize both open models and APIs at different stages of their workflows. “We aim to be the sole provider of both APIs and open models, offering the broadest range of capabilities for our community,” Banks stated.
Tris Warkentin, Director of Product Management for Google DeepMind, announced that the company will release comprehensive benchmarks evaluating Gemma alongside other models, which will be accessible on the OpenLLM leaderboards. “We are partnering with NVIDIA and Hugging Face to ensure that all public benchmarks have been run against these models,” he said, expressing pride in the transparent, community-focused approach taken during development.
Gemma is described as "responsible by design." Warkentin emphasized that these models have undergone extensive evaluations to ensure safety. The Google DeepMind blog post detailed that Gemma is aligned with AI Principles, incorporating automated techniques to filter out personal information from training datasets and utilizing reinforcement learning from human feedback (RLHF) to promote responsible behavior. Rigorous evaluations, including manual red teaming and automated adversarial testing, were conducted to assess the models' risk profiles.
Warkentin also stressed the importance of an open ecosystem in cultivating responsible AI. “We believe diverse perspectives from developers and researchers globally are essential for effective feedback and enhanced safety systems,” he stated. “Integrating this feedback and communication with the community is key to the value of this project.”