MarianMT / OPUS-MT

MarianMT is the neural machine translation framework that powers the OPUS-MT model collection from the University of Helsinki's language technology group. Marian itself is a fast C++ NMT toolkit, and the team trained and released over a thousand compact bilingual and multilingual checkpoints on the open OPUS parallel corpus, covering everything from major European languages to low-resource pairs. Each checkpoint is a small 6-layer encoder-decoder transformer that translates one direction and runs in real time on a laptop CPU.

These models became the default free, offline translation engine inside countless open-source apps, browser extensions, and research pipelines. The weights are CC-BY-4.0 and the Marian framework is MIT, so they are fully usable commercially. Quality is solid for high-resource pairs but clearly behind modern large models, and newer OPUS-MT work has largely been absorbed into the Tatoeba-Challenge releases.

No single published headline benchmark; quality tracked per language pair on Tatoeba-Challenge leaderboards (author-maintained)