From the source
Cohere researchers show that data mixing can make small reasoning models think in the user's language rather than defaulting to English, achieving over 93% in-language reasoning across 60 languages.
- The team built Tiny Aya L2-Thinker (3.35B parameters) on three data pillars: broad language coverage in reasoning data, a small share of multilingual non-reasoning data, and English reasoning data as a task-solving backbone.
- The model maintains accuracy close to English-only reasoning on math, language understanding, cultural reasoning, and open-ended writing, with the exception of competition-level math.
- Cohere releases the model weights and multilingual reasoning data, and says the approach generalizes to languages not seen in reasoning training examples.






