Google has unveiled Gemma, a family of open models designed to assist developers and researchers enganged in responsible AI development. Google said the Gemma models were created from the same research and technology used to create its Gemini models.

Described as a new generation of models, Gemma is a family of lightweight models developed by Google DeepMind and other teams across Google. Gemma model weights come in two sizes, Gemma 2B and Gemma 7B, both with pre-trained and instruction-tuned variants.

Gemma models can run directly on a developer laptop or desktop computer. Guides on Gemma open models are being provided on google.dev. Gemma outperforms larger models on key benchmarks while adhering to Google standards for safe, responsible outputs, Google said. Google has integrated Gemma with tools such as Hugging Face, MaxTex, Nvidia NeMo, and TensorRT-LLM.

Google also released a Responsible Generative AI Toolkit that provides guidance and tools for creating safer AI applications. The toolkit includes safety classification, debugging, and guidance. Also provided are reference implementations for inference and supervised fine-tuning across major frameworks including Keras, PyTorch, JAX, and TensorFlow/Keras.