mBART-50
Multilingual denoising-pretrained seq2seq model translating directly between any pair of 50 languages without an English pivot.
About
mBART-50 is Meta's multilingual extension of the BART denoising objective to machine translation. The model is pretrained by corrupting monolingual documents across 50 languages and learning to reconstruct them, then fine-tuned on many translation directions at once. Checkpoints ship in one-to-many, many-to-one, and many-to-many variants, and the key paper result is that multilingual fine-tuning improves zero-shot and low-resource translation well beyond single-pair training.
It became one of the most reused multilingual seq2seq backbones in both research and production prototypes, with straightforward Hugging Face integration. The Hugging Face weights carry a CC-BY-NC-4.0 license, so commercial use is restricted. With 12 encoder and 12 decoder layers at around 610M parameters, it runs comfortably on a single GPU.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Author-reported (paper): multilingual fine-tuning on the ML50 benchmark improves low-resource and zero-shot directions over bilingual baselines; no single headline FLORES score published
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: Meta FAIR paper (2020) confirmed via the Hugging Face model card; one of the most reused multilingual seq2seq checkpoints with one-to-many, many-to-one, and many-to-many variants. Strengths: 50-language direct translation, single-GPU footprint, proven fine-tuning base. Docked for: (1) CC-BY-NC-4.0 weights block commercial deployment; (2) superseded by NLLB-200 and M2M100 for language coverage. License: CC-BY-NC-4.0.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 9.0/10
- Maintenance · recency, releases, issue load
- 6.0/10
- Openness · license + docs
- 7.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · Meta FAIR 2020 paper; one-to-many, many-to-one, many-to-many checkpoints
- · ~610M params, runs on a single GPU
- · Widely used as a multilingual seq2seq fine-tuning base