Google has introduced Gemma, a collection of open models aimed at supporting developers and researchers committed to responsible AI development. Google states that Gemma models are built using the same research and technology behind its Gemini models.

Gemma represents a new era of models, developed by teams at Google DeepMind and other divisions within Google. These lightweight models, available in two sizes – Gemma 2B and Gemma 7B, come with pre-trained and fine-tuned variants.

One of Gemma’s notable features is its ability to run directly on developers’ laptops or desktop computers. Detailed guides on Gemma open models are accessible on google.dev. Despite their compact size, Gemma models outperform larger counterparts on essential benchmarks while adhering to Google’s safety and responsibility standards. Integration with tools like Hugging Face, MaxTex, Nvidia NeMo, and TensorRT-LLM further enhances Gemma’s utility.

In addition to Gemma, Google has released a Responsible Generative AI Toolkit, offering guidance and tools for creating AI applications with safety in mind. The toolkit includes features for safety classification, debugging, and comprehensive guidance. Reference implementations for inference and supervised fine-tuning across major frameworks such as Keras, PyTorch, JAX, and TensorFlow/Keras are also included.