Aya Vision

What Aya Vision is

Aya Vision is Cohere's first vision-language model: a model that takes both images and text as input and answers in text. Cohere's research lab released it on 4 March 2025 in two open-weight sizes, 8B and 32B parameters.^[1][3][4] Both sizes are published as downloadable weights rather than only behind an API, which is unusual for a company's first model of this kind.

How it works

The 32B model pairs the Aya Expanse 32B language model with a SigLIP2-patch14-384 vision encoder, and supports a context of 16k tokens.^[2] In a model built this way, the encoder turns an image into a sequence of vectors that the language model reads as if they were ordinary text tokens; the answer about the picture is then generated the same way any other text is continued.

Images are handled by cutting them into up to 12 tiles of 364x364 pixels plus a thumbnail. Each tile is encoded as 169 visual tokens, so a single image can occupy 2,197 tokens in total.^[2] Tiling is what lets a model read small details such as text in a screenshot, but it also means one picture eats a noticeable share of the 16k-token window — resolution and token budget have to be chosen together.

Languages

Aya Vision accepts inputs in 23 languages, including English, Japanese, Arabic and Hindi.^[1] Multilingual image understanding is harder than multilingual text, because the model has to connect what it sees to the vocabulary and grammar of each language rather than only to English. Aya Vision inherits its language model from the Aya line, the multilingual base it was built on.^[2]

Reported results

Cohere reported that the 8B model outperforms Llama 90B, roughly 11 times its size, and that the 32B model beats Qwen 72B, Llama 90B and Molmo 72B, with win rates up to 79% for the 8B and 72% for the 32B on AyaVisionBench and m-WildVision.^[1] The 32B model was evaluated specifically against Llama-3.2 90B Vision, Molmo 72B and Qwen2.5-VL 72B.^[2]

Win rates of this kind come from pairwise comparison: two models answer the same prompt and a judge picks the better answer. In these evaluations the judge was claude-3-7-sonnet-20250219.^[2] It is worth reading such figures with that in mind — they measure which answer is preferred, not whether an answer is correct, and a model acting as judge brings its own tendencies to the comparison.

Availability and licence

Aya Vision is released under CC BY-NC 4.0, which limits use to non-commercial purposes.^[1][2] The models are available on Cohere's website, on Hugging Face, on Kaggle, and through WhatsApp.^[1] Cohere also published AyaVisionBench, the evaluation set used in its comparisons, so that others can run the same tests.^[1]

The non-commercial clause is the practical catch for anyone hoping to build a product on the weights: open here means inspectable, runnable and testable, not freely usable in a commercial service.

Where it sits

Aya Vision is Cohere's entry into vision-language work, released with open weights at both sizes and accompanied by its own benchmark.^[1][4] It builds on the Aya Expanse 32B language model and adds a separate vision encoder on top,^[2] so its multilingual behaviour and its image handling come from different components joined together. That combination is the model's main claim: broad language coverage and image understanding in sizes small enough to run and inspect.