M2M100 1.2B (Translation)

M2M100 (Many-to-Many) 1.2B is a multilingual translation model from Meta that directly translates between any pair of 100 languages, without requiring English as a pivot language. This direct translation approach avoids the quality loss that occurs when translating through an intermediate language.

With 1.2 billion parameters, the model was trained on a massive multilingual corpus using novel data mining techniques to find parallel sentences across language pairs. It significantly improves translation quality for low-resource languages that previously lacked direct translation models.

M2M100 is a breakthrough in democratizing translation technology, enabling direct translation between language pairs like Arabic-to-French or Hindi-to-Japanese without English intermediary steps.