Gemma is a family of lightweight, state-of-the-art open models developed by Google DeepMind, based on the research and technology used to create Gemini models. These models demonstrate strong performance across academic benchmarks for language understanding, reasoning, and safety. Gemma offers two sizes of models (2 billion and 7 billion parameters) and provides both pre-trained and fine-tuned checkpoints. The models outperform similarly sized open models on 11 out of 18 text-based tasks. The release includes an open-source codebase for inference and serving, along with detailed evaluations of safety and responsibility aspects. Gemma's development emphasizes responsible deployment to improve safety, ensure equitable access, enable rigorous evaluation, and foster innovation. The models are trained on up to 6T tokens of text, using architectures, data, and training recipes inspired by Gemini. They achieve strong generalist capabilities in text domains and state-of-the-art understanding and reasoning skills. The paper discusses the model architecture, training infrastructure, pretraining and fine-tuning processes, and comprehensive evaluations across various benchmarks. It also addresses the responsible deployment of LLMs, including safety mitigations and ethical considerations.Gemma is a family of lightweight, state-of-the-art open models developed by Google DeepMind, based on the research and technology used to create Gemini models. These models demonstrate strong performance across academic benchmarks for language understanding, reasoning, and safety. Gemma offers two sizes of models (2 billion and 7 billion parameters) and provides both pre-trained and fine-tuned checkpoints. The models outperform similarly sized open models on 11 out of 18 text-based tasks. The release includes an open-source codebase for inference and serving, along with detailed evaluations of safety and responsibility aspects. Gemma's development emphasizes responsible deployment to improve safety, ensure equitable access, enable rigorous evaluation, and foster innovation. The models are trained on up to 6T tokens of text, using architectures, data, and training recipes inspired by Gemini. They achieve strong generalist capabilities in text domains and state-of-the-art understanding and reasoning skills. The paper discusses the model architecture, training infrastructure, pretraining and fine-tuning processes, and comprehensive evaluations across various benchmarks. It also addresses the responsible deployment of LLMs, including safety mitigations and ethical considerations.