Cohere releases Command R for enterprise RAG and tool use

Background

Companies that use AI for business tasks often need answers grounded in their own documents, and they need models that can call other software. Retrieval-augmented generation (RAG) is a common way to supply that grounding: the model retrieves relevant documents first and then writes its answer from them. Cohere, a Canadian enterprise AI company, designed Command R for this kind of work.

What happened

On 11 March 2024 Cohere released Command R^[1][2]. The model was built for RAG and tool use at production scale, which means it was intended for real business systems rather than only for research demonstrations^[1].

Key details

The model has 35 billion parameters and a context window of up to 128k tokens^[3][2]. A context window is the amount of text the model can consider in a single request.

Cohere said the model performs strongly in 10 languages, including English, French, Japanese, Korean, Arabic and Chinese^[2].

Grounded answers include citations to the source documents, and the model can call tools in single or multiple steps^[3]. A citation lets a reader trace each claim back to the document it came from.

Cohere For AI released the model's weights on Hugging Face for research use. Commercial use requires a licence^[2][3]. The weights are the trained parameters, so the release allowed researchers to download and run the model themselves.

The model was offered through Cohere's hosted API at lower pricing than Cohere's earlier models^[2].

Impact

The release offered business users a model built around grounded answers and tool use, at lower pricing than Cohere's earlier models^[2]. For readers following the Command line, this was the point at which the family first focused on those two features. The next steps in the line are covered in the Command R entry.

Limits

The benchmarks and comparisons published with the release come from Cohere, and they measure specific tasks. Because commercial use requires a licence, the open weights cannot be used commercially without one.