Command R+
Last reviewed
Sources
11 citations
Review status
Source-backed
Revision
v1 · 1,678 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
11 citations
Review status
Source-backed
Revision
v1 · 1,678 words
Add missing citations, update stale details, or suggest a clearer explanation.
Command R+ is a large language model developed by Cohere, the Toronto- and San Francisco-based enterprise artificial intelligence company, and released on April 4, 2024.[1][2] A 104-billion-parameter model with a 128,000-token context window, Command R+ was the larger and more capable member of Cohere's Command R series at launch, optimized for retrieval-augmented generation (RAG) with verifiable in-line citations, multi-step tool use (function calling), and multilingual enterprise workloads.[3][4] Positioned as a cost-effective, scalable alternative to GPT-4-class systems for business use, Command R+ helped define the retrieval-optimized enterprise-model class and was the first model in Cohere's lineup made available on Microsoft Azure.[1][2] Cohere released the model's weights publicly on Hugging Face under a non-commercial license, with commercial access offered through its own API and major cloud platforms.[3]
Command R+ targets enterprise use cases that require accurate, grounded, and auditable outputs: long-document question answering, summarization, knowledge-base search, and agentic workflows that chain together external tools and APIs. Its two defining features are RAG with citations, in which the model generates answers from supplied document snippets and annotates each claim with the source it came from, and multi-step tool use, in which the model can call several tools in sequence and feed the result of one call into the next.[3][4] Cohere paired these capabilities with a 128,000-token context window and optimization for ten key business languages, framing the model as the production-grade tier of its Command R family.[3][4]
The model launched alongside, and as the larger sibling of, Command R, a 35-billion-parameter model aimed at the same RAG and tool-use scenarios at lower cost.[5] Cohere later extended the line with Command R7B, a compact 7-billion-parameter model released in December 2024, and with Command A in March 2025, a successor designed to match the capabilities of the Command R+ class while running on substantially less hardware.[6]
The following table summarizes the principal characteristics of Command R+. Figures reflect Cohere's published model documentation and Hugging Face model card.
| Attribute | Detail |
|---|---|
| Developer | Cohere and Cohere Labs (formerly Cohere For AI) |
| Initial release | April 4, 2024 |
| Refreshed version | August 30, 2024 (command-r-plus-08-2024) |
| Parameters | 104 billion |
| Context window | 128,000 tokens |
| Architecture | Auto-regressive, optimized transformer; supervised fine-tuning plus preference training |
| Modality | Text (input and output) |
| Key capabilities | RAG with in-line citations, multi-step tool use, multilingual generation |
| Optimized languages | English, French, Spanish, Italian, German, Brazilian Portuguese, Japanese, Korean, Arabic, Simplified Chinese (10) |
| Additional pre-training languages | 13 more (e.g., Russian, Polish, Turkish, Vietnamese, Dutch, Czech, Indonesian, Hindi, Hebrew) |
| Open weights | Hugging Face, under CC-BY-NC with an acceptable-use addendum |
| Commercial access | Cohere API, Microsoft Azure, Amazon Bedrock, Oracle Cloud Infrastructure |
Cohere was founded in 2019 by Aidan Gomez, Nick Frosst, and Ivan Zhang, with Gomez being one of the co-authors of the 2017 transformer paper "Attention Is All You Need." Unlike rivals focused on consumer chatbots, Cohere positioned itself as an enterprise-first provider, emphasizing data privacy, deployment flexibility, and models tuned for business tasks rather than general-purpose assistants.[7]
The Command R series formalized that strategy. Cohere introduced Command R in March 2024 as a scalable model optimized for RAG and tool use, then unveiled Command R+ on April 4, 2024, describing it as its most capable model to that point and the production-scale tier built for "enterprise-grade workloads."[1][5] Command R+ debuted first on Microsoft Azure before reaching Cohere's own API and other cloud platforms, a sequencing Cohere highlighted as part of its push to meet enterprises where their data and infrastructure already lived.[1][2]
On August 30, 2024, Cohere refreshed both models, releasing command-r-08-2024 and command-r-plus-08-2024.[8] According to Cohere, the updated Command R+ delivered roughly 50 percent higher throughput and 25 percent lower latency than the original release while keeping the same hardware footprint, alongside improvements in coding, mathematics, reasoning, multilingual RAG, instruction following, and structured-data handling.[8][9] Cohere also reported that the refreshed 35-billion-parameter Command R had become competitive with the original Command R+ on many tasks.[8][9]
Command R+ is an auto-regressive language model that uses an optimized transformer architecture, refined after pre-training through supervised fine-tuning and preference training to align its outputs with human preferences for helpfulness and safety.[3] Cohere built the model around three enterprise-oriented capabilities.
The model is trained to perform grounded generation: given a query and a set of retrieved document snippets, it produces an answer drawn from those sources and inserts citations indicating exactly which document supports each statement.[3][4] In the open-weights release, these grounding spans are expressed with markup so that applications can surface the underlying sources to users, a feature Cohere frames as central to delivering "faithful and verifiable" responses and reducing hallucination in regulated and high-stakes settings.[3]
Command R+ supports tool use, also called function calling, in which the model decides when to invoke an external tool and constructs the appropriate call. Beyond single calls, it is trained for zero-shot multi-step tool use, meaning it can chain several tools across multiple steps and use the output of one call as input to the next, which enables the construction of simple autonomous agents.[3][4] The August 2024 refresh specifically improved the model's judgment about which tool to use in a given context and whether a tool is needed at all.[8]
Command R+ is optimized for ten key business languages: English, French, Spanish, Italian, German, Brazilian Portuguese, Japanese, Korean, Arabic, and Simplified Chinese, and was pre-trained on a total of 23 languages.[3][4] Cohere designed an accompanying tokenizer to be efficient on non-Latin scripts; in one comparison the company noted that OpenAI's tokenizer produced about 1.67 times as many tokens for Japanese text, which directly affects per-token API costs.[10]
Cohere positioned Command R+ as a cost-effective, scalable competitor to GPT-4-class models, with its relative strengths concentrated in RAG and tool use rather than raw general knowledge. The benchmark claims below are Cohere's own evaluations reported at launch and should be read as vendor-reported figures.
On tool-use evaluations, Cohere reported that Command R+ outperformed GPT-4 Turbo, and on multi-hop RAG question-answering benchmarks it described the model as highly competitive with, or ahead of, similarly priced models such as Claude 3 Sonnet and Mistral Large.[4][10] Independent community evaluation lent some external support to the launch: on the public LMSYS Chatbot Arena leaderboard, Command R+ was reported in April 2024 as the top-ranked open-weights model at the time, placing competitively with some GPT-4 versions on human preference voting.[10] As with all such results, scores depend on prompt format, shot count, and evaluation date, and later models from Cohere and competitors surpassed these figures.
At its April 2024 launch, Command R+ was priced at 3.00 US dollars per million input tokens and 15.00 US dollars per million output tokens through Cohere's API.[11] With the August 2024 refresh, Cohere lowered list pricing for command-r-plus-08-2024 to 2.50 US dollars per million input tokens and 10.00 US dollars per million output tokens, undercutting GPT-4-class pricing while emphasizing the model's RAG and agentic strengths.[9][11]
Cohere made Command R+ available through two distinct channels. The model weights were released publicly on Hugging Face by Cohere Labs (then Cohere For AI) under a Creative Commons Attribution-NonCommercial (CC-BY-NC) license, supplemented by an acceptable-use addendum and Cohere Labs' Acceptable Use Policy; this permits research and non-commercial experimentation but not commercial deployment of the open weights.[3] For production use, Cohere offered managed commercial access through its own API and through major cloud platforms, with Command R+ launching first on Microsoft Azure and subsequently becoming available on Amazon Bedrock and Oracle Cloud Infrastructure, among others.[1][2][8] This dual approach, open weights for the research community alongside commercial API and cloud distribution, became a recurring pattern across Cohere's Command lineup.
Command R+ was an early and influential example of an enterprise large language model designed specifically around retrieval and tool use rather than open-ended conversation. By coupling a 128,000-token context window with grounded generation, source citations, and multi-step function calling, it gave organizations a model whose outputs could be traced back to their underlying documents, a property valued in finance, healthcare, legal, and other settings where unsupported answers carry real cost.[3][4] Its positioning as a cheaper, RAG-focused alternative to GPT-4-class models, backed by vendor benchmarks emphasizing tool use and multilingual retrieval, exemplified Cohere's broader strategy of competing on enterprise fit and total cost of ownership rather than on frontier general intelligence.[7][10]
The model also helped popularize the retrieval-optimized enterprise tier as a distinct product category, one that Cohere itself iterated on rapidly: the August 2024 refresh improved efficiency and capability at lower prices, Command R7B extended the approach to compact deployments in December 2024, and Command A succeeded the Command R+ class in March 2025 by delivering comparable performance on far less hardware.[6][8] Within that trajectory, Command R+ stands as the model that established Cohere's RAG-and-tools formula at production scale.