ATLAS: Practical scaling laws for multilingual models
Google Research
Researchers published ATLAS, a set of scaling laws for training multilingual AI models across 400+ languages, based on 774 training runs spanning 10M–8B parameters. The study quantifies cross-lingual transfer effects between 1,400 language pairs and shows that supporting twice as many languages requires increasing model size by 1.18x and training data by 1.66x. The work provides practical guidance for developers building models for non-English speakers on optimal language mixtures, model sizes, and whether to pre-train from scratch or fine-tune from multilingual checkpoints.
Why it matters
Generative AI
