Cohere says its Tiny Aya L2-Thinker model reasons in a user’s language more than 93% of the time across 60 languages, with accuracy close to English reasoning outside competition-level maths. If that result holds up in practice, it could make AI reasoning easier for multilingual users to follow and check, rather than leaving them with an answer in their language and the thinking in English.
Cohere Watch analysis
What happened
The Daily AI Digest reports that Cohere used a new training-data approach in Tiny Aya L2-Thinker to encourage reasoning in the language used by the person asking. It says the model also showed some ability to reason in languages for which it had not seen reasoning examples.
The digest attributes the findings to Cohere, but does not provide the underlying report, evaluation method or a breakdown by language. The headline figure is therefore a promising claim, not enough on its own to show how reliably the approach works across all 60 languages.
Why it matters
Reasoning people can actually read is more useful than a polished answer whose route to the conclusion remains opaque. For organisations serving multilingual teams or customers, the reported result could make it easier for people to inspect how a model reached an answer, and to catch errors without switching languages.
That matters especially because a strong average can conceal uneven performance. Sixty languages is a wide field; the useful question is whether the gains reach speakers of less-resourced languages too, rather than concentrating where training data is plentiful.
Our read
This is a meaningful direction for multilingual AI, and the reported exception for competition-level maths is worth keeping in view. Cohere’s result deserves attention, but a headline percentage is not a substitute for language-by-language results and a clear account of how “reasoning” was assessed. Until those details are available, take it as an encouraging claim, not a settled benchmark.
What to watch
- Whether Cohere publishes the evaluation method and results for individual languages.
- How the approach performs in languages with less training data.
- Whether the claimed multilingual reasoning ability carries over to real user tasks.
Discussion spark: Would you trust an AI’s reasoning more if it explained itself in your language, or would you need language-by-language evidence before relying on it?
Sources and evidence
not affiliated with or endorsed by Cohere