RELEASE · MODELS · #1586
Cohere releases Tiny Aya L2-Thinker (3.35B) that reasons in users' languages
Cohere introduced Tiny Aya L2-Thinker, a 3.35B multilingual reasoning model (built on the Tiny Aya 32K-context base) that produces reasoning traces in the prompt language over 93% of the time across 60 languages. Cohere says the model achieves this L2 (in-language) reasoning via a data-mixing training strategy, preserves near-English accuracy across varied tasks, and the team is releasing the model weights and multilingual reasoning data.
KEY POINTS
- Cohere introduced Tiny Aya L2-Thinker, a 3.35B multilingual reasoning model (built on the Tiny Aya 32K-context base) that produces reasoning traces in the prompt language over 93% of the time across 60 languages.
- Cohere says the model achieves this L2 (in-language) reasoning via a data-mixing training strategy, preserves near-English accuracy across varied tasks, and the team is releasing the model weights and multilingual reasoning data.
- Releasing a small, open-weight model and multilingual reasoning data that reasons in users' languages lowers barriers for non-English inspection, evaluation, and further research on multilingual reasoning.
WHY IT MATTERS
Releasing a small, open-weight model and multilingual reasoning data that reasons in users' languages lowers barriers for non-English inspection, evaluation, and further research on multilingual reasoning.