Cohere Command A Vision

Command A Vision, published on 31 July 2025, was the first Cohere model able to read images. Its brief is narrow and deliberately unglamorous: charts, graphs, diagrams, tables, optical character recognition, document question answering and object detection. This is the visual work that happens inside an insurance claim or a quarterly report, not image generation or creative interpretation. The model accepts up to 20 images in a single request, which matters for document workflows where a single case file is a stack of scanned pages rather than one photograph. Its context window is 128,000 tokens with 8,000 tokens of output, and it officially supports English, Portuguese, Italian, French, German and Spanish — a shorter list than the 23 languages of its text-only siblings, which is worth noting before assuming the whole Command A family behaves identically. Structurally it is Command A with sight added: 111.87 billion parameters against 111.06 billion for the text model, the difference being roughly 810 million parameters of vision tower bolted on top. Cohere kept the same API surface as the rest of the Command line, so adding image input to an existing integration is a model-name change rather than a rewrite. The weights are on Hugging Face under CC BY-NC 4.0 — research and inspection are allowed, commercial deployment is not.

▸Videos

No videos yet.

▸Reviews