Cohere Labs releases Tiny Aya L2-Thinker, a 3.35B model that reasons in 44 languages
lmoroney · x · 2026-10-09
Reasoning models typically think in English even when prompted in Hindi or Swahili, making their reasoning traces unreadable to many users.
Cohere Labs released Tiny Aya L2-Thinker on Hugging Face: a 3.35B open-weights research model trained to reason in the prompt's language, covering 44 languages plus English with 32K context. Cohere reports in-language reasoning over 93% of the time across 60 languages, with accuracy within two to three points of an English-reasoning twin on five of six benchmarks; competition-level math is the exception.
The recipe centers on data mix: 1.7M English reasoning examples, 5,000 translated reasoning examples per language, plus some multilingual data without reasoning traces. Licensed CC-BY-NC, so commercial use requires talking to Cohere.
More from Models
- Claude Haiku 5.5 jumps 257 points to 1587 on WebDev Arena leaderboard — arena · 2026-10-09
- Private visual reasoning bench eyebench saturated by Astra at 97%, author rules out data contamination — adonis_singh · 2026-10-09
- Don't crown Anthropic yet: six months ago it throttled hard, and the pendulum will swing back — ryunuck · 2026-10-09
- Haiku-5.5 reportedly hits #7 on EyeBench-v3, 76x cheaper than fable-5.1-max — adonis_singh · 2026-10-09
- ProximalHQ's training setup lifts Qwen3.8-27B to 37.2% Pass@1 on DeepSWE, up 8.4 points — aryaman2020 · 2026-10-09
- Grok moderation debate: the architecture is the culprit, not the model — xtel9 · 2026-10-09