Cohere Labs' Tiny Aya L2-Thinker hits 90%+ in-language reasoning in 60 languages at just 3.35B
Cohere_Labs · x · 2026-09-11
Cohere Labs releases Tiny Aya L2-Thinker, a 3.35B-parameter model with 32K context trained to reason in the user's own language rather than defaulting to English.
- Data-centric SFT recipe mixing multilingual reasoning, multilingual non-reasoning, and English reasoning data
- Achieves above 90% (93% in the paper) in-language reasoning rate across 60 languages on 6 benchmarks: math, commonsense, cultural reasoning, instruction following, open-ended generation
- Ships open weights plus multilingual reasoning data covering 44 languages beyond English — the largest coverage the team is aware of
- Key finding: generalizing L2 reasoning to held-out languages needs broad language coverage, readily available non-reasoning data, and a strong English reasoning backbone
- Paper: arXiv:2609.10445
More from Models
- DeepSeek V4.1 Flash weights out: 552B multimodal, 1M context, SGLang day-0 support — zhaoran_wang · 2026-09-11
- CursorBench 4.0 rolls out with harder, longer-horizon coding tasks, scores drop — StringChaos · 2026-09-11
- Developer: Astra is a good model — jxnlco · 2026-09-11
- New Model Hits Opus-Level Benchmarks at Wild Efficiency, RL Infra Details Emerge — nrehiew_ · 2026-09-11
- DeepSeek V4.1 Flash notes: how obsessing over KV cache compression yields a hyper-efficient frontier model — nrehiew_ · 2026-09-11
- Model with reasoning off finds bugs Opus 5 missed — cephaloform · 2026-09-11