Command R+

What Command R+ is

Command R+ is a large language model built by Cohere and announced on 4 April 2024. Cohere presented it as its most capable model optimised for retrieval-augmented generation, and the launch post describes it as a state-of-the-art RAG-optimised model designed for enterprise-grade workloads^[1]. It was available first on Microsoft Azure^[1][4]. The model has 104 billion parameters and a context window of 128k tokens^[2].

Retrieval-augmented generation is the pattern Command R+ was tuned for: instead of answering only from what the model memorised during training, a system first retrieves relevant documents from a search index or database and passes them to the model together with the question, so the answer can be grounded in those documents and traced back to them. The context window matters here because it sets how much retrieved material can be handed to the model at once.

Documents, citations and tool use

By default, the model card says, Command R+ produces grounded responses by first predicting which documents are relevant^[2]. It offers two citation modes, "accurate" and "fast"; in the fast mode, a step in the tokenizer generates the answer directly with the grounding spans inside it^[2].

The model also adds stronger multi-step tool use: it can chain several tool calls, moving through action, observation and reflection steps before producing an answer^[2]. That lets the model gather information in stages — call a tool, look at what comes back, decide what to do next — rather than committing to a single call.

On language coverage, Command R+ is optimised for the same 10 languages as Command R, with 13 additional languages covered during pre-training^[2].

Two ways to get it

Microsoft added Command R+ to the Azure AI model catalog as part of its Models as a Service offering, and the model works with Azure AI Search and Azure AI Content Safety^[4].

Separately, the weights were released on Hugging Face by Cohere For AI under CC-BY-NC 4.0, a non-commercial licence^[2]. Cohere Labs describes the release as an open-weights research release of a 104 billion parameter model^[2]. The two channels serve different purposes: the Azure route is a managed service inside Microsoft's cloud, while the Hugging Face release lets researchers download and study the weights under terms that do not permit commercial use.

Updates and succession

An August 2024 update, command-r-plus-08-2024, delivered roughly 50% higher throughput and 25% lower latency on the same hardware, along with better tool-use decisions and new safety modes^[3]. Its training data runs through February 2023^[3]. The model card on Hugging Face points to the newer c4ai-command-r-plus-08-2024 version as the one available^[2].

In March 2025 Cohere succeeded it with Command A, which the company said had 150% higher throughput than Command R+ 08-2024^[3].

Where it sits

Command R+ was Cohere's strongest RAG-optimised model at launch^[1], shipped both as a managed service on Azure^[4] and as downloadable non-commercial weights^[2], refreshed in August 2024^[3], and replaced by Command A in March 2025^[3]. Its design emphasis is that retrieval, citation and staged tool use are handled by the model itself rather than left entirely to the surrounding application^[2].