Cohere's Tiny Aya L2-Thinker: 3.35B model hits 93% in-language reasoning across 60 languages
Cohere_Labs · x · 2026-09-16
Cohere released Tiny Aya L2-Thinker, a 3.35B model with 32K context trained to reason in the user's language instead of defaulting to English.
Key results
- 90%+ (93% per abstract) in-language reasoning rate across 60 languages on 6 benchmarks (math, commonsense, instruction following, open-ended generation, cultural reasoning)
- Beats models 2-7x its size on in-language reasoning rate and task accuracy for data-scarce languages
Method: data-centric SFT mixing multilingual reasoning, multilingual non-reasoning, and English reasoning data. Generalization to held-out languages comes from broad language coverage, readily available non-reasoning data, and a solid English reasoning backbone.
Takeaway: reasoning is a language-agnostic behavior transferable via careful data mixing, without per-language reasoning supervision. Weights and multilingual reasoning data covering 44 non-English languages are open-sourced.
More from Research
- Sloman's 1981 paper: emotions are inevitable in machines managing multiple motives — yeastsplainer · 2026-09-16
- fal's H3 Max: RL post-training cuts 5s video generation from 120s to under 3s — isidentical · 2026-09-16
- Agent theory researcher quits, calling the work "emotionally taxing" — IgorCarron · 2026-09-16
- Science LLM benchmarks have flawed answers; fixing them significantly raises model scores — Profanion · 2026-09-16
- Aaronson hears AI companies have cracked longstanding TCS open problems, sitting on major announcements — scaling01 · 2026-09-16
- 'Cognitive Surrender' paper released, link shared by economist Paul Novosad — paulnovosad · 2026-09-16