CodeGen provides open-source models trained on TPU-v4 for program synthesis, spanning CodeGen1 and CodeGen2 releases with parameter sizes from 350M to 16B. The models accept natural language prompts and generate executable code, supporting multi-turn program synthesis interactions. CodeGen2 introduced strong infill sampling capability, while CodeGen2.5 achieved performance competitive with 16B models using only 7B parameters. The models are available via Hugging Face and have been published at ICLR 2023.