# Cohere releases Tiny Aya L2-Thinker with 93% in-language reasoning

> The new multilingual model reasons in the user's language for 60 languages, keeping accuracy almost unchanged.

Oossa · 2026-10-06 · https://oossa.com/en/cohere-releases-tiny-aya-l2-thinker-with-93-in-language-reasoning

Cohere announced Tiny Aya L2-Thinker, a multilingual reasoning model that produces reasoning traces in the user's language. It achieves over 93% in-language reasoning across 60 languages while accuracy drops only 2‑3 points on most benchmarks. The model builds on the Tiny Aya base and uses a mix of English reasoning data, translated multilingual examples, and non‑reasoning data. Cohere made the model and training data available on Hugging Face.

## The facts

- Published Oct 06 2026
- In-language reasoning rate > 93% across 60 languages

## Why it matters

Non‑English speakers can read and verify AI reasoning steps in their own language, improving transparency.

## Sources & references

1. [Multilingual Bridges: How Data Mixing Unlocks In-Language ReasoningWhen a reasoning model gets an answer wrong, you’d read its reasoning to find out why. But what if that reasoning is in a language yo](https://cohere.com/blog/building-multilingual-bridges) – Cohere, 2026-10-06

Last updated: 2026-10-06