Cohere’s Command A+ brings sovereign AI within reach of any enterprise with a server room

Canadian AI company Cohere has unveiled Command A+, a large language model designed for complex reasoning, document processing, and autonomous workflows. Carl Franzen reports for VentureBeat that the model is available for free on Hugging Face under an Apache 2.0 license, making it the first Cohere model that anyone can use, modify, and commercialize without restrictions.

Cohere CEO Aidan Gomez confirmed that this is the first time the company has released a model under such a permissive open-source license. Previous Cohere models, including Command R and Command R+, were available for research but prohibited commercial use without a paid license.

Efficient by design

Command A+ has 218 billion parameters in total, but only 25 billion are active at any given time. This so-called Sparse Mixture-of-Experts architecture routes each query only to the most relevant parts of the model, keeping resource use low. The result is a model that can run on a single NVIDIA Blackwell B200 GPU or just two NVIDIA H100 GPUs, making it far more accessible than competing proprietary models estimated to have trillions of parameters.

Cohere also applied heavy compression to the model through a technique called quantization, reducing its memory footprint while preserving accuracy. According to the company, this approach achieves up to 63% faster output speed compared to its predecessor.

Built for enterprise use

• Native citation generation links every factual claim to its source, reducing the risk of undetected errors
• Support for 48 languages, with improved efficiency for Arabic, Japanese, and Korean
• A 128,000-token context window for processing long documents, images, and charts
• Direct integration with databases, APIs, and search tools for autonomous workflows

On key benchmarks, the model showed substantial gains over its predecessor. Its score on a complex telecom reasoning test jumped from 37% to 85%. On a difficult mathematics test, it reached 90%, up from 57%.

Because the Apache 2.0 license allows deployment on private servers or air-gapped networks, companies in regulated industries can run the model entirely within their own infrastructure, with no dependency on Cohere’s services.