The objective of work package 3 “Database and MT Infrastructure” is to develop and implement the MMT machine translation architecture and its core components, in a way to meet the scalability and big data requirements of the project. It also includes aspects of research and innovation in order to achieve machine translation training in real time and to improve translation quality through the exploitation of context information. This report is the third and last deliverable of WP3. It describes the current status of the technology developed by the consortium during the third year of the project. The described version has been also integrated in the final MMT public release (D4.4). This report is arranged as follows: Section 3 introduces the new neural MMT architecture that has been developed during the third year of the project. Section 4 describes the new memory component of the architecture, while Section 5 discusses the deployed neural adaptive MT engine. Section 6 reviews the pre- and post-processing modules of the architecture. Section 7 reviews the final version of MMT with respect to its initial requirements set in D1.2. Finally, Section 7 presents results of the new neural MMT architecture on an industrial use case. The report is completed by an Appendix that provides all main technical details of the neural adaptive MT approach implemented in MMT, which were published at the 2017 Conference on Machine Translation (WMT17)1.