Cohere releases Command A Vision

Background

Command A is Cohere's language model, and Command A Vision was built on top of it ^[1].

What happened

On 31 July 2025 Cohere released Command A Vision, a vision-language model built on Command A ^[1]. The model has 112 billion parameters and uses a SigLIP2 vision encoder to read images ^[1]. Its accompanying blog post carries the same date ^[1].

Key details

The model reads images, charts and documents and answers in text ^[2][1]. It can handle images of up to about 3 megapixels, which are split into 512 by 512 tiles before processing ^[1].

The open weights were posted on Hugging Face under a CC-BY-NC licence, which allows non-commercial use only ^[1]. Commercial customers therefore go through Cohere itself.

Impact

The release put a vision model whose weights could be downloaded from Hugging Face into public hands, although the non-commercial licence limited what could be done with it ^[1]. Oracle later added the model, together with Command A Reasoning, to its OCI Generative AI service, which meant business customers could reach it through a major cloud platform ^[2].