Gemma 2 is a powerful tool built for developers and researchers, designed to seamlessly integrate into their workflows. Available under the commercially-friendly Gemma license, Gemma 2 allows users to share and commercialize their innovations easily. With broad framework compatibility, Gemma 2 can be used with major AI frameworks such as Hugging Face Transformers, and JAX, PyTorch, and TensorFlow via native Keras 3.0, vLLM, Gemma.cpp, Llama.cpp, and Ollama. The software is optimized with NVIDIA TensorRT-LLM for NVIDIA-accelerated infrastructure or as an NVIDIA NIM inference microservice, with optimization for NVIDIA’s NeMo in the works.
Effortless deployment of Gemma 2 will be available for Google Cloud customers starting next month through Vertex AI. The Gemma Cookbook offers practical examples and recipes to help users build their applications and fine-tune Gemma 2 models for specific tasks. From retrieval-augmented generation to exploring usage with preferred tools, the Gemma Cookbook guides users every step of the way.
Committed to responsible AI development, the Gemma team provides resources to ensure developers and researchers can build and deploy AI responsibly. The Responsible Generative AI Toolkit includes the recently open-sourced LLM Comparator for in-depth evaluation of language models. The toolkit also includes a Python library to run comparative evaluations with models and data, with visualization capabilities in the app. The team is actively working on open-sourcing text watermarking technology, SynthID, for Gemma models.
When training Gemma 2, thorough safety processes were followed, including filtering pre-training data and rigorous testing and evaluation against a comprehensive set of metrics to identify and mitigate potential biases and risks. Results of these safety processes are published on public benchmarks related to safety and representational harms to ensure transparency and accountability.