Multilingual Bridges: How Data Mixing Unlocks In-Language Reasoning
Cohere, Tuesday, October 6th, 2026
Cohere shows how data mixing lets a 3.35B-parameter model reason in the user's language over 93% of the time across 60 languages.
Cohere argues that truly multilingual AI should reason in the user's language, since English-only reasoning can lose nuance and keeps non-English speakers from reading reasoning traces.
Its research achieves in-language reasoning through data mixing built on broader language coverage in reasoning data, a small share of multilingual non-reasoning data, and English reasoning data as the task-solving backbone.
The resulting Tiny Aya L2-Thinker (3.35B) reasons in the prompt language over 93% of the time across 60 languages, with minimal accuracy loss, solid low-resource performance and fewer reasoning tokens than comparable models.