Command R

What Command R is

Command R is a large language model from Cohere, announced on 11 March 2024 and built for retrieval-augmented generation (RAG) and tool use at production scale ^[1][3]. It has 35 billion parameters and a 128,000-token context window ^[2][4]. Cohere's documentation describes the model as having a long 128,000-token context length ^[4], and Cohere Labs describes it on Hugging Face as a research release of a 35 billion parameter generative model ^[2]. The launch post is dated 11 March 2024 and written by Cohere co-founder and CEO Aidan Gomez ^[1].

Retrieval-augmented generation is a way of getting a model to answer from documents instead of only from what it absorbed during training: relevant passages are fetched from a corpus, placed in the prompt, and the model is expected to work from them. That pattern is common in enterprise settings, where answers have to be traceable to an internal document rather than produced from memory. Command R was designed around it rather than treated as an add-on.

Retrieval, citation and tools

The model supports grounded generation, meaning it can produce answers that cite the document snippets it was given ^[2]. It is also designed to work with Cohere's Embed and Rerank retrieval models, which handle the searching and reordering of candidate passages before they reach the model ^[3].

On the tool side, Command R supports single-step tool use, often called function calling, in which the model decides to invoke an external function and then answers using the result. It also supports multi-step tool use, the pattern needed for agents that chain several calls together before producing a final answer ^[2]. Together, retrieval and tool use cover two common ways of connecting a language model to information and actions outside itself.

Languages and context

The model is optimised for 10 languages: English, French, Spanish, Italian, German, Brazilian Portuguese, Japanese, Korean, Simplified Chinese and Arabic. A further 13 languages are included in pre-training ^[2]. The 128,000-token context window allows a large amount of retrieved material to be placed in a single prompt, which matters when the passages come from long documents ^[2][4].

Availability and licence

Command R is available through Cohere's hosted API ^[3][2]. Cohere For AI released the weights on Hugging Face under the CC-BY-NC 4.0 licence, which is non-commercial, together with Cohere's acceptable-use policy ^[3][2]. The Hugging Face model card describes the release as a non-quantized version of Command-R, with a 4-bit quantised version kept in a separate repository ^[2].

Evaluation at launch

Cohere's launch materials compared the model with Mixtral in human preference evaluations on enterprise RAG tasks, and reported end-to-end accuracy on the Natural Questions, TriviaQA and HotpotQA benchmarks ^[3]. End-to-end accuracy here refers to the score on the whole retrieval-plus-answer pipeline, not just the generation step in isolation.

Updated version and place in the line

An updated version, command-r-08-2024, was released in August 2024, with better tool-use decisions and instruction following, roughly 50% higher throughput and 20% lower latency, and half the hardware footprint needed to serve it ^[4].

Command R was followed in April 2024 by the larger Command R+ ^[4]. Cohere now recommends Command A for most use cases ^[4], which makes Command R a step in the line rather than its current endpoint: a model whose defining choices were a long context, retrieval as a first-class feature, and tool use built in for production deployments.