Cohere has unveiled the Tiny Aya model family, a collection of small (~3.3 billion parameter) multilingual language models catering to more than 70 languages. This suite is designed to include languages that are often less represented in mainstream models. The lineup consists of a base model and a general-purpose ‘Global’ model, along with three region-specific variants: Earth, Fire, and Water. These specialized models are tailored for the linguistic needs of West Asia, Africa, Europe, South Asia, and the Asia-Pacific regions, respectively, by integrating region-specific SFT checkpoints.
Cohere’s innovation extends to creating a custom tokenizer optimized for multilingual performance, which stands as a competitor to Gemma 3 and Qwen 3 tokenizers. Notably small, these models are efficient enough to run on mobile devices, allowing for the development of language assistants in areas not covered by larger models.
The training process employed a pool of 64 H100 GPUs, showcasing the models’ adaptability to modest computing resources. This efficiency makes them suitable for offline applications on typical devices like laptops, empowering regions with limited internet access to harness AI for translation and other applications. Cohere has also released accompanying multilingual training datasets and benchmarks to facilitate further fine-tuning.
For deployment, Tiny Aya models are readily accessible on platforms such as HuggingFace, Kaggle, and Ollama, enabling users to integrate and utilize these models locally.

Collections

Cohere has released the Tiny Aya model family, a set of compact (~3.3B parameter) multilingual language models supporting over 70 languages, including many underrepresented ones. The lineup includes a base model, a general-purpose Global model, and three region-specific variants (Earth, Fire, Water) targeting West Asia, Africa, Europe, South Asia, and Asia-Pacific. A custom multilingual tokenizer competes with Gemma 3 and Qwen 3 tokenizers. Trained on 64 H100 GPUs, the models are efficient enough to run on mobile devices and laptops, making them suitable for offline use in regions with limited internet access. Multilingual training datasets and benchmarks are also released, and the models are available on HuggingFace, Kaggle, and Ollama.

Cohere Unveils Tiny Aya: Compact Multilingual Models for Diverse Languages