A multilingual assistant can accept a question in Arabic, Hindi or Swahili and return a polished answer in the same language while producing its intermediate reasoning in English. For a user who cannot inspect English, the visible answer is local but the deliberation remains inaccessible. Cohere calls the alternative in-language, or L2, reasoning.
Its Tiny Aya L2-Thinker is a 3.35-billion-parameter model trained to reason in the prompt language. Cohere reports an in-language reasoning rate above 93% across 60 languages, with relatively small accuracy losses on five of six evaluation groups and more noticeable weakness on competition-level mathematics.[1][2]
What Cohere reports
| Measure | Value | Context |
|---|---|---|
| Model size | 3.35B parameters | A compact multilingual model rather than a frontier-scale system.Source 1Source 3 |
| Evaluation coverage | 60 languages | Tests include mathematics, cultural reasoning, instruction following and open-ended writing.Source 1Source 2 |
| In-language reasoning | Above 93% | Share of traces identified as being in the prompt language.Source 1 |
| Training languages | 45 | The model shows transfer to some languages without direct reasoning examples.Source 1 |
| Context | 32K | The work builds on the Tiny Aya base model with a 32,000-token context window.Source 1 |
Answer language and reasoning language are separate capabilities
Translation at the end can hide the language in which a model decomposed the problem, interpreted cultural context or made an error. In-language reasoning aims to preserve nuance earlier in the process and lets a user examine the steps without switching languages.
Cohere’s data-mixing result is notable because it does not require a complete reasoning corpus for every language. English reasoning data supplies much of the task-solving backbone. A comparatively small multilingual reasoning set teaches the model which language to use, and multilingual question-answer data helps the behaviour transfer to additional languages.[1][2]
Three training components and their roles
| Data component | What it teaches | Observed role |
|---|---|---|
| English reasoning | How to solve step-by-step tasks | Strong task backbone but mostly English tracesSource 1 |
| Multilingual reasoning | How reasoning looks in target languages | Raises in-language reasoning sharply from the English-only baselineSource 1 |
| Multilingual non-reasoning | Questions and answers without traces | Improves accuracy and transfer, including languages without reasoning examplesSource 1 |
A visible trace is useful evidence, not a certificate
A reasoning trace can help a user spot a mistranslation, cultural assumption or repeated loop. It can also be incomplete, post-hoc or only loosely connected to the mechanism that produced the answer. The trace should be treated as an inspectable explanation generated by the model, not a guaranteed transcript of an internal process.
Cohere itself examines repetition and long unproductive traces. Some comparison models spend many thousands of tokens circling the same steps before failing to answer. Tiny Aya’s shorter and less repetitive traces are promising, but efficiency should be judged alongside correctness and task completion rather than as a virtue on its own.[1]
The test should include language-specific meaning
Mathematics is valuable for measuring reasoning, but it often reduces the importance of language. A strong multilingual evaluation also needs culturally grounded questions, ambiguous instructions, local conventions, formality, idiom and open-ended writing. A model may translate numbers perfectly while flattening the social meaning of a request.
- Compare final-answer accuracy and reasoning language separately.
- Evaluate high- and low-resource languages rather than reporting one average.
- Ask native speakers to assess nuance, clarity and inappropriate cultural assumptions.
- Measure whether traces help users find and correct errors.
- Record repetition, unfinished answers and reasoning-token cost.
In-language reasoning expands the definition of multilingual quality. The question is no longer only whether the final answer can be translated. It is whether people can follow the model’s work, challenge its assumptions and use it without English becoming an invisible gate. Cohere’s result suggests that broader access may be possible through smarter data design rather than model size alone.
Sources & further reading
- Multilingual Bridges: How data mixing unlocks in-language reasoning Cohere · 6 Oct 2026 · Accessed 9 Oct 2026
- Building Multilingual Bridges: Data Mixing as the Pillar of Generalization for In-Language Reasoning arXiv · 9 Sep 2026 · Accessed 9 Oct 2026
- Tiny Aya Cohere Documentation · Accessed 9 Oct 2026