Multilingual model
A model trained on many languages so it can understand and write in each, and often translate between them.
Most large models are multilingual to some degree, because the web is. But English dominates most training sets, often making up the large majority of tokens, so quality drops in other languages, especially those with less text online. Builders improve coverage by adding data in more languages, enlarging tokenizers and fine-tuning on translated instructions. BLOOM in 2022 was trained on 46 natural languages including several Indian languages, and Cohere's Aya project gathered instruction data in over 100 languages with volunteers around the world.