Cohere launches Command R+, first on Microsoft Azure

Background

By early 2024 many businesses wanted AI systems that could answer questions from their own documents and carry out tasks with other software. Retrieval-augmented generation (RAG) meets the first need by having the model look up relevant documents before it answers. Cohere, a Canadian enterprise AI company, had released Command R in March 2024 for this purpose, and Command R+ followed a few weeks later.

What happened

On 4 April 2024 Cohere introduced Command R+, a model it described as optimised for RAG in enterprise settings^[1]. The model was made available first on Microsoft Azure, through the Azure AI model catalogue's Models as a Service offering^[1][2].

Key details

The model has 104 billion parameters and a context window of 128k tokens^[3]. A context window is the amount of text the model can consider in a single request.

It supports multi-step tool use for agent workflows, meaning the model can carry out a sequence of actions, one after another, in order to complete a task^[3]. It also gives grounded answers with citations, which let readers check where each statement came from^[3].

Cohere For AI published the weights on Hugging Face under a non-commercial CC-BY-NC licence^[3]. Non-commercial means the weights may be used for research and other non-business purposes, while commercial use needs separate permission.

Microsoft said that the model works with Azure AI Search and Azure AI Content Safety^[2]. Azure AI Search is a search service that developers use to build retrieval systems. Azure AI Content Safety is a service for screening content for harm.

Impact

The launch gave enterprise customers a model on a major cloud platform, which is one way businesses can use a model without running their own infrastructure. The announcement described the model's tool use and citations^[1][3].

Limits

The figures and descriptions in this entry come from Cohere and Microsoft's announcements. They describe what the model is designed to do, not how it performs on every task. Because the weights are non-commercial, businesses that want to use them directly need a separate arrangement.