Aya Vision 32B extends the Aya Expanse family into the multimodal space, combining a 32 billion parameter architecture with vision capability to handle text and images together in 23 languages. According to Cohere's documentation, the model is positioned as a state-of-the-art system on language, text, and image benchmarks, supporting tasks such as image captioning, visual question answering, text generation, and translation drawn from either text or images into coherent written output. The third-party comparison surfaces a DocVQA multilingual score around 76.3 percent, giving one concrete sense of how the model performs on document-style visual question answering, while the underlying design is oriented toward unifying communication across cultures and geographies rather than specializing in any single narrow domain.
Because the weights are openly released under the CohereLabs organization on Hugging Face, Aya Vision 32B is well suited for practitioners who want to self-host, fine-tune, or experiment with a multilingual vision-language model without depending solely on a closed API. The practical sweet spot is cross-lingual scenarios where images need to be described, queried, or translated for speakers of languages that mainstream vision models often overlook, including multilingual captioning pipelines, accessible image descriptions, and bilingual visual question answering. At the same time, the model is not framed as a coding or general-purpose agent specialist, so teams seeking heavy agentic function calling or code generation will likely pair it with other tools rather than rely on it alone.
