Gemma is a family of lightweight, open-source language models developed by Google DeepMind, based on research and technology from the Gemini models. The models are trained on up to 6T tokens of text and come in two sizes: 2B and 7B parameters. They demonstrate strong performance across academic benchmarks for language understanding, reasoning, and safety. Both pretrained and fine-tuned checkpoints are released, with Gemma outperforming similarly sized open models on 11 out of 18 text-based tasks. Comprehensive evaluations of safety and responsibility are provided, along with detailed model development descriptions. The responsible release of LLMs is critical for improving safety and enabling future innovations. Gemma is trained on up to 6T tokens of text, using architectures, data, and training recipes inspired by the Gemini model family. It achieves strong generalist capabilities in text domains and state-of-the-art understanding and reasoning skills. The models are available for efficient deployment on GPU and TPU, and for CPU and on-device applications. They are designed to address different computational constraints and applications. The models are evaluated on a wide range of quantitative and qualitative benchmarks, including automated and human evaluations. Gemma models demonstrate strong performance on domains such as question answering, commonsense reasoning, mathematics, and coding. They outperform other models on several benchmarks, including MMLU and MBPP. The models are also evaluated for memorization, with results showing low rates of memorization compared to other models. Personal and sensitive data memorization is not observed. Gemma is developed with a focus on safety and responsible deployment. The models are evaluated for safety and responsible use, with a structured approach to responsible development and deployment. The release of Gemma is intended to enable further AI safety research and community innovation, while also addressing the risks associated with open models. The models are available for a wide range of applications, including science, education, and the arts. They are designed to support a variety of developer needs, with two model sizes to optimally support different environments. The release of Gemma is intended to reduce economic and technical barriers for developers and to enable the development of beneficial applications. The models are also available for further research and development, with a detailed model card and a Generative AI Responsible Toolkit provided to support responsible AI development.Gemma is a family of lightweight, open-source language models developed by Google DeepMind, based on research and technology from the Gemini models. The models are trained on up to 6T tokens of text and come in two sizes: 2B and 7B parameters. They demonstrate strong performance across academic benchmarks for language understanding, reasoning, and safety. Both pretrained and fine-tuned checkpoints are released, with Gemma outperforming similarly sized open models on 11 out of 18 text-based tasks. Comprehensive evaluations of safety and responsibility are provided, along with detailed model development descriptions. The responsible release of LLMs is critical for improving safety and enabling future innovations. Gemma is trained on up to 6T tokens of text, using architectures, data, and training recipes inspired by the Gemini model family. It achieves strong generalist capabilities in text domains and state-of-the-art understanding and reasoning skills. The models are available for efficient deployment on GPU and TPU, and for CPU and on-device applications. They are designed to address different computational constraints and applications. The models are evaluated on a wide range of quantitative and qualitative benchmarks, including automated and human evaluations. Gemma models demonstrate strong performance on domains such as question answering, commonsense reasoning, mathematics, and coding. They outperform other models on several benchmarks, including MMLU and MBPP. The models are also evaluated for memorization, with results showing low rates of memorization compared to other models. Personal and sensitive data memorization is not observed. Gemma is developed with a focus on safety and responsible deployment. The models are evaluated for safety and responsible use, with a structured approach to responsible development and deployment. The release of Gemma is intended to enable further AI safety research and community innovation, while also addressing the risks associated with open models. The models are available for a wide range of applications, including science, education, and the arts. They are designed to support a variety of developer needs, with two model sizes to optimally support different environments. The release of Gemma is intended to reduce economic and technical barriers for developers and to enable the development of beneficial applications. The models are also available for further research and development, with a detailed model card and a Generative AI Responsible Toolkit provided to support responsible AI development.