Cohere releases Aya Vision, its first vision model

Background

Aya Vision was built on Aya Expanse 32B, a multilingual open-weight language model from Cohere's research lab, and added a SigLIP2 vision encoder so that it could handle images ^[2].

What happened

On 4 March 2025 Cohere's research lab released Aya Vision in 8B and 32B open-weight sizes ^[1]. The models understand images and text across 23 languages ^[1]. It was Cohere's first vision model ^[1].

Key details

Cohere said the 8B model beat Llama 90B, and the 32B model beat Qwen 72B, Llama 90B and Molmo 72B, on multilingual image tasks ^[1]. The weights are licensed under CC BY-NC 4.0, which means they can be used for non-commercial purposes only ^[1].

Users could try the model through WhatsApp, as well as through Hugging Face and Kaggle ^[1].

Impact

The release brought Cohere's multilingual approach into vision, with open weights that researchers could test on their own machines. Because of the non-commercial licence, companies that wanted to build products on the model needed to make separate arrangements ^[1].