CodeGemma is a collection of open-source code models based on Google's Gemma models. The models are trained on a large amount of code and natural language data, achieving state-of-the-art performance in code completion and generation tasks. Three variants are released: CodeGemma 7B pretrained (PT) and instruction-tuned (IT) models, and a specialized 2B model for code infilling and open-ended generation. The 2B model is particularly fast, making it suitable for latency-sensitive applications. CodeGemma models excel in mathematical reasoning and code generation, outperforming other open models in these areas. They also maintain strong natural language understanding. The models are designed for practical use in latency-constrained environments, such as IDEs and local environments. CodeGemma is evaluated on various benchmarks, including HumanEval, BabelCode, and mathematical reasoning tasks, demonstrating its effectiveness in real-world coding scenarios. The models are also compared with other code models in the same size class, showing superior performance in mathematical reasoning. The paper provides detailed instructions for using the models, including formatting for code completion tasks and inference recommendations. The models are open-source and available for use in a wide range of applications.CodeGemma is a collection of open-source code models based on Google's Gemma models. The models are trained on a large amount of code and natural language data, achieving state-of-the-art performance in code completion and generation tasks. Three variants are released: CodeGemma 7B pretrained (PT) and instruction-tuned (IT) models, and a specialized 2B model for code infilling and open-ended generation. The 2B model is particularly fast, making it suitable for latency-sensitive applications. CodeGemma models excel in mathematical reasoning and code generation, outperforming other open models in these areas. They also maintain strong natural language understanding. The models are designed for practical use in latency-constrained environments, such as IDEs and local environments. CodeGemma is evaluated on various benchmarks, including HumanEval, BabelCode, and mathematical reasoning tasks, demonstrating its effectiveness in real-world coding scenarios. The models are also compared with other code models in the same size class, showing superior performance in mathematical reasoning. The paper provides detailed instructions for using the models, including formatting for code completion tasks and inference recommendations. The models are open-source and available for use in a wide range of applications.