tiny-aya-en-thinker

Tiny Aya En-Thinker is an open-weights research checkpoint from Cohere Labs built on the Tiny Aya family (Cohere2 autoregressive transformer, 3.35 billion parameters). It is supervised fine-tuned for multilingual reasoning with explicit chain-of-thought: internal reasoning traces are produced in English while final answers follow the user’s prompt language. The model supports roughly 44 reasoning languages plus broader coverage via multilingual non-reasoning instruction data, and exposes dual modes via /think and /no_think in the chat template.

It targets mathematics, science, and general reasoning, instruction following, and multilingual open-ended generation when English thinking with localized answers is desired. Cohere positions it as the English-reasoning twin of Tiny Aya L2-Thinker (same base and training mixture, but L2-Thinker keeps reasoning in the prompt language). Intended uses include multilingual conversational assistants, research on cross-lingual reasoning traces, and low-resource deployment at small scale; use Tiny Aya L2-Thinker when reasoning must stay in the user’s language.

The checkpoint ships under CC-BY-NC 4.0 with Cohere Labs’ acceptable-use policy and a gated Hugging Face download. Context length is 32K tokens (input plus output). Primary public evaluation in the accompanying paper reports multilingual reasoning suites; on English MGSM (GSM8K-style grade-school math), En-Thinker scores 92.8% accuracy with near-100% English reasoning rate.