Cohere Labs releases Tiny Aya L2-Thinker for in-language reasoning across 60 languages

Cohere Labs has released Tiny Aya L2-Thinker, a 3.35B open-weight research model designed to reason in the language of the user’s prompt instead of defaulting to English reasoning traces. The work builds on the Tiny Aya base model with a 32K context length and studies what Cohere calls L2, or in-language, reasoning across 60 languages. The publisher reports that the resulting model produces reasoning in the prompt language more than 93% of the time across its evaluation set while keeping most task accuracy close to an English-reasoning counterpart.

The central contribution is the data mixture rather than a language-specific architecture. Cohere Labs combines about 1.7 million English reasoning examples with multilingual reasoning examples translated into 44 languages and a third pool of multilingual question-and-answer data without reasoning traces. In the publisher’s ablations, the multilingual reasoning slice sharply raises the rate of in-language reasoning, while the non-reasoning multilingual data helps generalization to languages that do not have dedicated reasoning supervision. Cohere also calls out competition-level mathematics as a harder case where the accuracy trade-off is more visible.

Cohere Labs says it is releasing both the model weights and multilingual reasoning data on Hugging Face. For Open Model Weights, this is a model-release candidate as well as a training-method signal: the canonical record should separately verify the exact repository artifacts, license, context and lineage, while the reported multilingual benchmark and reasoning-rate results remain attributed research claims. The release is especially relevant to the evidence layer because the training recipe is described in enough detail to record source-backed methodology without turning benchmark comparisons into a universal quality ranking.