Aya Vision 32B

Aya Vision 32B is the larger of the two vision-language models Cohere Labs published on 3 March 2025 and the only one carried into the company's production catalogue, where it appears as c4ai-aya-vision-32b with a 16,000-token context window and a 4,000-token output limit. It reads images alongside text and answers in the 23 languages of the Aya line, Polish included, with image captioning and visual question answering named as the primary tasks. The repository puts the model at 33,137,896,688 parameters. Set against the 32,296,476,672 of Aya Expanse 32B, the vision half accounts for roughly 841 million - about 2.5% of the whole, and the same construction used in the 8B variant, where the proportion is nearly three times higher. Licensing follows the family: CC BY-NC 4.0 and the acceptable use policy on the downloadable weights, no commercial grant, gate before download, and separate commercial terms if the model is rented through Cohere instead. Downloads are modest for a model of this size - 706 in the thirty days to 18 August 2026, against 226 likes - which is the usual pattern for weights that need two or more datacentre GPUs to run and cannot be used to earn anything.

▸Videos

No videos yet.

▸Reviews