The paper introduces OpenNMT, an open-source toolkit for neural machine translation (NMT) that prioritizes efficiency, modularity, and extensibility. The toolkit aims to support research in model architectures, feature representations, and source modalities while maintaining competitive performance and reasonable training requirements. Key features include memory sharing for GPU-based training, multi-GPU support, and various deployment options for different run-time environments. The system is implemented in Lua/Torch and has been extended to support Python/PyTorch. Case studies demonstrate the toolkit's ability to handle factored neural translation and different types of attention mechanisms. Additionally, OpenNMT supports multiple modalities, such as image-to-text and speech-to-text, and includes tools for tokenization, word embeddings, and visualization. Benchmarks show improved performance compared to existing systems, making OpenNMT a valuable resource for both academic and industrial NMT research.The paper introduces OpenNMT, an open-source toolkit for neural machine translation (NMT) that prioritizes efficiency, modularity, and extensibility. The toolkit aims to support research in model architectures, feature representations, and source modalities while maintaining competitive performance and reasonable training requirements. Key features include memory sharing for GPU-based training, multi-GPU support, and various deployment options for different run-time environments. The system is implemented in Lua/Torch and has been extended to support Python/PyTorch. Case studies demonstrate the toolkit's ability to handle factored neural translation and different types of attention mechanisms. Additionally, OpenNMT supports multiple modalities, such as image-to-text and speech-to-text, and includes tools for tokenization, word embeddings, and visualization. Benchmarks show improved performance compared to existing systems, making OpenNMT a valuable resource for both academic and industrial NMT research.