Omnilingual MT: Machine Translation for 1,600 Languages
Summary
Meta AI unveils Omnilingual MT (OMT), a machine translation system that supports 1,600 languages by combining public multilingual data with new datasets and advanced evaluation artifacts. The work presents two model approaches (decoder-only OMT-LLaMA and encoder–decoder OMT-NLLB) and reports that 1B–8B parameter models can match or exceed a 70B baseline, significantly expanding generation capabilities for long-tail languages; open datasets and models are freely available.