Cohere's Command A Vision Beats GPT-4.1 on Documents Running on Just Two A100s
Cohere releases Command A Vision, a 112B open-weights multimodal model that beats GPT-4.1 on document, chart, and OCR tasks while running on just two A100 GPUs.
Takeaways
- Open-weights multimodal model: Cohere releases Command A Vision, a 112B vision-language model under a CC-BY-NC license.
- Beats GPT-4.1 overall: Scores 83.1% average across 9 benchmarks vs. GPT-4.1's 78.6%, leading on ChartQA, DocVQA, OCRBench, and TextVQA.
- Runs on 2x A100s: Deployable on just two A100 GPUs, or a single H100 with 4-bit quantization -- unusually efficient for a 112B model.
- Built for documents, not general vision: Excels at OCR, chart parsing, and PDF Q&A; trails GPT-4.1 by ~9.5 points on MMMU general reasoning.
- 128K context, 20 images per request: Supports long multi-image conversations; does not support tool/function calling.
- LLaVA-style architecture: Combines SigLIP2 vision encoder with the Command A 111B text tower via an MLP adapter, trained with RLHF using Contrastive Policy Gradient.
Cohere has released Command A Vision, its first multimodal model and the latest addition to the Command A family. At 112 billion parameters, it is built to do one thing exceptionally well: understand the kinds of images that actually show up in enterprise workflows -- scanned documents, dense charts, tables, and annotated diagrams. The weights are open, the license is CC-BY-NC, and it runs on surprisingly modest hardware for its size.
The enterprise vision gap
Most large language models these days are multimodal, but enterprises generally use more graphical documents such as charts and PDFs, so extracting information from these unstructured data sources often proves difficult. That is the gap Command A Vision is designed to close. It empowers businesses to automate tedious tasks, unlock valuable insights from visual data, and make highly accurate, data-driven decisions through document OCR and image analysis.
With Deep Research on the rise, the importance of bringing in models capable of reading, analyzing and even downloading unstructured data has grown. Cohere is betting that enterprises want a self-hostable option for this -- one that does not require sending sensitive documents to a closed API.
How it is built
Command A Vision integrates a SigLIP2 vision encoder with the existing 111B Command A text tower to produce a model that processes documents, charts, and tables at competitive accuracy with far lower deployment costs than closed API alternatives. The architecture follows the LLaVA pattern -- a well-established approach where a vision encoder feeds into a language model through a lightweight connector layer (called an MLP adapter).
Command A Vision is an enterprise-focused multimodal large language model that excels at understanding and interpreting visual data (images, charts, documents) alongside text. Here is how images are processed under the hood:
- Each image is split into up to 12 tiles of 512x512 pixels, plus one global thumbnail tile