fairseq2

It provides production-quality implementations of sequence-to-sequence architectures, language model training infrastructure, and multilingual NLP components that underpin Meta's large-scale translation, speech, and language model research, while exposing clean APIs that external researchers can use to build on.

The library reimplements core components with better abstractions than the original fairseq: more composable model definitions, improved multi-GPU and multi-node distributed training support, native integration with modern PyTorch features like torch.compile and mixed-precision training, and cleaner dataset and data loading infrastructure.

fairseq2 supports the training patterns used in Meta's NLLB (No Language Left Behind) multilingual translation models, SeamlessM4T speech translation, and related large-scale multilingual systems.

NLP researchers training sequence models at scale, speech processing teams building multilingual translation and recognition systems, and practitioners using Meta's pretrained models (available through fairseq2's model hub) use this library as a foundation.

The transition from the original fairseq reflects Meta AI's accumulated experience running large training runs and the software engineering lessons learned from maintaining a widely-used research codebasemaking fairseq2 more maintainable and extensible for both Meta's internal research and the broader open-source NLP community.