command-a-vision-07-2025

Command A Vision 07-2025 is Cohere Labs' open-weights 112-billion-parameter vision-language model built on the Command A text backbone and a SigLIP2-patch16-512 vision encoder connected via an LLaVA-style multimodal adapter. It targets enterprise image understanding—document OCR, chart and diagram analysis, table extraction, and multilingual visual QA—while retaining strong text capabilities from Command A, including RAG-oriented workflows. Images are tiled into up to twelve 512×512 patches plus a global thumbnail (up to 3,328 visual tokens per image), with support for up to 20 images per request and recommended resolutions up to 2048×1536.

On nine standard vision benchmarks, Command A Vision averages 83.1%, outperforming GPT-4.1 (78.6%), Llama 4 Maverick (80.5%), and Pixtral Large (76.8%). It leads strongly on document-centric tasks—DocVQA 95.9%, ChartQA 90.9%, OCRBench 86.9%—and scores 73.5% on MathVista. General multimodal reasoning on MMMU (chain-of-thought) is 65.3%, below GPT-4.1's 74.8%. Training follows three stages: vision-language alignment, supervised fine-tuning with model merging, and RLHF-style post-training for enterprise safety and quality.

Released under CC-BY-NC for research use (commercial licensing via Cohere), the model supports a 128K-token context window, six vision languages (English, French, German, Italian, Portuguese, Spanish), and configurable safety modes (contextual or strict). It is available on Hugging Face, the Cohere API, and Oracle Cloud Infrastructure, and is designed for efficient private deployment on as few as two GPUs.