MADLAD-400

MADLAD-400 is Google Research's multilingual machine translation project: a manually audited, 3-trillion-token monolingual dataset spanning 419 languages, plus baseline models trained on it — including a 10.7B-parameter translation model trained on 250 billion tokens covering over 450 languages. Released alongside the 2023 paper 'MADLAD-400: A Multilingual And Document-Level Large Audited Dataset', it was explicitly built to serve low-resource languages that mainstream MT systems ignore.

The released variants — madlad400-3b-mt, madlad400-7b-mt (plus a back-translation variant), and madlad400-10b-mt — are plain encoder-decoder T5-style models usable directly in Hugging Face Transformers with no fine-tuning. The team reported the 10.7B model competitive with significantly larger systems across domains, and its strength is breadth: hundreds of languages where commercial APIs either don't exist or perform poorly.

Honest caveats: it is a research artifact, not a production service — no hosted API, modest English quality versus modern LLMs, and the 2023-era architecture predates today's best MT. Check the Hugging Face model card for the exact license before commercial use. For researchers and NGOs working in low-resource languages, it remains one of the most accessible open MT resources ever released.