Cohere says its 3.35-billion-parameter Tiny Aya L2-Thinker reasons in the language of its prompt more than 93% of the time across 60 languages. The company’s research suggests multilingual reasoning can be made more accessible without a large accuracy penalty on most of its tests, though competition-level maths remains a notable weakness.
Cohere Watch analysis
What happened
Cohere describes the model and its training approach in a research post published on 6 October. The team combined English reasoning examples, a smaller amount of translated reasoning data in 44 languages, and multilingual question-and-answer data without reasoning traces. Cohere says that mix helped the model carry in-language reasoning into languages it had not seen reasoning examples for.
In the company’s evaluations across 60 languages and six benchmarks, the model reasoned in the prompt’s language more than 93% of the time. Cohere reports accuracy drops of no more than two to three points on five benchmarks compared with its English-reasoning counterpart, with a slight improvement on its open-ended writing benchmark. The exception was PolyMath, a competition-level maths test, where accuracy fell more noticeably. Cohere says it is releasing the model weights and multilingual reasoning data.
Key findings
- Reasoning in the prompt’s language
Cohere reports an in-language reasoning rate above 93% across 60 languages. - A data-mixing recipe
English reasoning data supplies problem-solving examples; translated reasoning and multilingual non-reasoning data help broaden language coverage. - One clear limitation
The company reports a more noticeable accuracy decline on competition-level maths. - Materials for further work
Cohere says it is releasing the model weights and multilingual reasoning data.
Why it matters
A translated final answer is not the same as reasoning a reader can follow. If the intermediate reasoning is in a language they understand, people may have a better chance of checking a model’s steps and spotting mistakes. Cohere also argues that reasoning in the user’s language could help models draw on cultural and linguistic knowledge expressed most naturally there.
The practical point is the training recipe, not a claim that this smaller model has settled multilingual AI. These are Cohere’s own evaluations, and the maths result shows the trade-offs have not vanished. Still, the reported performance makes the case that in-language reasoning need not be built from scratch, language by language.
Our read
This is a useful research result with a reader-facing consequence: a model that can show its working in more languages may be easier for more people to scrutinise. The reported breadth is promising; the next test is whether independent evaluations reproduce it, especially on harder tasks and lower-resource languages. The weights and data give other researchers something concrete to examine, rather than asking them to take a tidy chart on trust.
What to watch
- Whether independent evaluations confirm the reported in-language reasoning rates and accuracy.
- How performance holds up across individual languages, particularly those with less training data.
- Whether later training improves the competition-level maths result.
Discussion spark: Would you trust a model’s reasoning more if it showed its working in your language, even if that came with a small accuracy trade-off?
Sources and evidence
- Source update (6 October 2026, 12:00 UTC)
not affiliated with or endorsed by Cohere