Command A Vision

What it is

Command A Vision is the first vision-language model in Cohere's Command family, released in July 2025; the accompanying blog post is dated 31 July 2025 ^[1]. It accepts text and images as input and produces text as output ^[1]. The model card frames it as a 112-billion-parameter model optimised for enterprise image understanding tasks ^[1].

How it is built

The model has 112 billion parameters in total. Architecturally it joins a SigLIP2-patch16-512 vision encoder to a Command A language model through a multimodal adapter, so the language model receives visual information in a form it can process alongside text ^[1].

Images are handled as tiles: each 512x512 tile is encoded as 256 visual tokens, and a single image can use up to 12 tiles plus a thumbnail, for a maximum of 3,328 tokens ^[1]. This tile budget is the practical way the model controls how much of the context an image occupies. The model supports a context window of 128k tokens, although the configuration published on Hugging Face defaults to 32k ^[1].

It works in English, Portuguese, Italian, French, German and Spanish ^[1].

What it is meant for

The emphasis on document and chart reading is stated by both the model card and by Oracle, which describes Command A Vision as a multimodal model for understanding images, charts and documents and added it to OCI Generative AI ^[2]. Oracle says the model is well suited to business vision-language applications, including detailed document examination and extracting key insights from images ^[2]. The model identifier used there is cohere.command-a-vision-07-2025, and Oracle gives its release date as July 2025 ^[2].

How it is distributed

The weights are published on Hugging Face under CC-BY-NC, a non-commercial licence; commercial use is arranged through Cohere ^[1]. Alongside the open weights, the model is offered as a managed service: Oracle's release notes give 21 January 2026 as the release date on which Command A Reasoning and Command A Vision became available in OCI Generative AI ^[2].

Where it sits in the line

Command A Vision extends the Command A language model into multimodal input rather than starting from a separate architecture, which is why the parameter count and the name track the existing model ^[1]. Its design choices — tiles of 512x512 pixels, a per-image cap of roughly 3.3 thousand visual tokens, and a 128k context — describe a model built around long documents and detailed images rather than short captions ^[1]. For enterprise use it arrives in two forms at once, as downloadable weights and as a cloud service ^[1][2].