Cohere Command A Vision

Command A Vision is Cohere's first multimodal model, released July 31, 2025 (model ID command-a-vision-07-2025). Built on the 112B-parameter Command A, it adds image understanding to Cohere's enterprise-focused lineup: up to 20 images per request, 128K-token context, 8K max output, and the same Command API interface. It's tuned for business document work — chart and graph analysis, table extraction, OCR with natural-language querying, and multilingual image text (English, French, German, Spanish, Italian, Portuguese). Cohere reports it surpassing GPT-4.1, Llama 4 Maverick, and Mistral Medium on standard vision benchmarks.

The honest picture: the weights are open but under CC-BY-NC — research and evaluation only, commercial use needs a license or the paid API. Tool use isn't supported with this model, which limits agentic workflows. And those benchmark wins are Cohere's own claims; independent verification is thinner than for the big multimodal APIs. It's a solid enterprise document-vision option, not a general creative multimodal model.