Marian NMT

High-performance neural machine translation framework written in pure C++ for maximum inference speed and training efficiency.

Built in pure C++ by teams at Microsoft and the University of Edinburgh, Marian NMT is the engine behind Microsoft Translator. It prioritizes raw performance (utilizing Intel MKL and AVX for CPU inference) and scales across multiple GPUs for rapid training. The framework supports modern Transformer architectures and includes specialized features like teacher-student training and beam search. It remains a top choice for production environments where latency and throughput are critical metrics.

https://marian-nmt.github.io/