Command R+

Command R+ is a large language model developed by Cohere, released in April 2024 as the flagship of the Command R generation and distinguished by its 104-billion-parameter architecture, 128,000-token context window, and optimization for retrieval-augmented generation, multi-step tool use, and multilingual enterprise workloads. It is available through the Cohere API, major cloud platforms including Amazon Bedrock and Microsoft Azure, and as open weights under a CC-BY-NC license via Hugging Face for non-commercial research. As of April 2026, Command A (March 2025) has replaced Command R+ as Cohere's primary commercial flagship, offering double the context window and 150% higher throughput, but Command R+ remains in active production and retains relevance as an open-weights reference model for the enterprise RAG use case Cohere built its identity around.

At a glance

• Lab: Cohere
• Released: April 4, 2024 (initial); August 2024 (refreshed Command R+ 08-2024 version)
• Modality: Text
• Open weights: Yes, for non-commercial research. CC-BY-NC 4.0 license with Cohere's Acceptable Use Policy. Available at CohereLabs/c4ai-command-r-plus on Hugging Face.
• Context window: 128,000 tokens
• Pricing: $2.50 per million input tokens / $10.00 per million output tokens (Cohere API; Command R+ 08-2024)
• Distribution channels: Cohere API, Amazon Bedrock, Microsoft Azure AI Foundry, Oracle Cloud Infrastructure (OCI), Hugging Face (open weights)

Origins

The Command family at Cohere predates the R generation by several years. The original Command model launched in November 2021 as a general-purpose instruction-following language model sold exclusively to enterprise customers via API. From the start, Cohere's positioning separated it from the consumer-AI trajectory of OpenAI and Anthropic: no chatbot product, no consumer subscription, no engagement with the public assistant market. The founding thesis was that frontier-capability language models would become enterprise infrastructure, and the early Command models were built for that channel specifically.

Command R arrived in March 2024, marking the transition to a new model generation designed explicitly for retrieval-augmented generation at enterprise scale. The "R" designation signals RAG as a first-class design goal: the model was trained with grounding and citation as core output behaviors, not post-hoc additions. Command R launched on Hugging Face under a CC-BY-NC license at 35 billion parameters, a deliberate contrast to the closed model lines of its competitors.

Command R+ arrived on April 4, 2024, debuting on Microsoft Azure before becoming broadly available through the Cohere API and, later, Amazon Bedrock. At 104 billion parameters, Command R+ represented the high-capability end of the Command R generation. The launch positioning framed it as competitive with GPT-4 Turbo and Claude 3 Opus on enterprise-relevant tasks, particularly RAG quality and tool use chaining, though independent benchmark comparisons offered a more differentiated picture than the launch claims suggested.

Cohere refreshed Command R+ in August 2024, releasing the command-r-plus-08-2024 model variant with roughly 50% higher throughput and 25% lower latency compared to the April 2024 version, while keeping hardware requirements identical. The August refresh also included improvements to math, coding, and structured output quality. Open weights for the August variant were published to Hugging Face under the same CC-BY-NC license as the original.

Command A, released in March 2025, succeeded Command R+ as Cohere's commercial flagship. Command A is a 111-billion-parameter model with a 256,000-token context window and 150% higher throughput than Command R+ 08-2024. Command R+ has not been deprecated and continues to serve active production workloads and research use through the Cohere API and Hugging Face, but the enterprise-priority role has moved to Command A.

Capabilities

Command R+'s defining capability is retrieval-augmented generation with inline citations. The model is trained to generate responses grounded in retrieved documents and to attribute specific claims to source passages with structured citation markers. This behavior is part of the model's output format rather than a prompting workaround: RAG with citations is a first-class output mode, activated through Cohere's API using the documents parameter alongside the user query. For enterprise deployments where auditability and factual grounding are requirements, this is a practical difference from models that cite sources only when prompted to do so.

Multi-step tool use is the second structural capability. Command R+ can plan and execute sequences of tool calls across multiple turns, combining outputs from one step as inputs to subsequent calls without requiring the application to manage each step explicitly. This enables agent-style workflows where the model coordinates API calls, database lookups, and intermediate computations in a single generation chain.

Multilingual coverage spans 23 languages total. Command R+ is performance-optimized for 10: English, French, Spanish, Italian, German, Brazilian Portuguese, Japanese, Korean, Simplified Chinese, and Arabic. Pre-training data covers 13 additional languages: Russian, Polish, Turkish, Vietnamese, Dutch, Czech, Indonesian, Ukrainian, Romanian, Greek, Hindi, Hebrew, and Persian. Cohere has cited multilingual RAG -- retrieving and generating in the same non-English language without routing through English -- as a differentiated capability versus English-dominant models.

Structured outputs, safety modes, and customizable system prompts complete the API feature set. Safety modes allow deployments to calibrate content filtering for the appropriate enterprise context without application-layer post-processing. The model supports JSON and other structured output formats for integration with downstream data pipelines.

The 128,000-token context window accommodates long documents such as legal contracts, research reports, and financial filings without chunking. At launch in April 2024, 128K context was near the frontier; by April 2026, with Command A at 256K and Claude Opus 4.7 at longer contexts, Command R+ sits in the mid-range.

Benchmarks and standing

Command R+'s benchmark profile reflects its enterprise RAG focus rather than general-capability competition with frontier models. At the April 2024 launch, Cohere cited performance competitive with GPT-4 Turbo on RAG-specific and tool-use evaluations; independent evals offered a more conditional picture, with Command R+ leading on RAG citation quality for supported deployment patterns and trailing on general reasoning benchmarks.

On the Hugging Face Open LLM Leaderboard at time of the original open-weights release, Command R+ scored approximately 74.6 average across the leaderboard suite, placing ahead of DBRX Instruct (74.5) and Mixtral 8x7B Instruct (72.7) and establishing it as a competitive open-weights model for the 100B-parameter class. These evaluations date to mid-2024; the leaderboard has since been revised and these direct comparisons no longer appear in the current scoring.

For RAG-specific evaluation frameworks such as RAGAS (which measures faithfulness, answer relevance, and context precision), Command R+'s citation-aware training provides a structural advantage: the model outputs citations in a format that can be verified against source documents without requiring additional processing. This is less a benchmark result than an architectural feature that determines where Command R+ fits in deployment stacks.

On general-purpose evaluations as of April 2026, Command R+ has been surpassed by the leading closed-source frontier models. GPT-5.5, Claude Opus 4.7, and Gemini 3.1 Pro score in the mid-to-high 50s on the Artificial Analysis Intelligence Index composite; Command R+ and its successor Command A are not prominently featured in the April 2026 rankings, reflecting Cohere's departure from direct benchmark competition with the frontier consumer labs.

Benchmark figures are point-in-time and shift as labs release updates and evaluation frameworks evolve.

Access and pricing

The primary production access path is the Cohere API, available at dashboard.cohere.com. Command R+ 08-2024 is priced at $2.50 per million input tokens and $10.00 per million output tokens, placing it at the same tier as Command A (March 2025) in Cohere's pricing structure. A free trial tier with rate limits is available for evaluation. Enterprise agreements with volume pricing and on-premises deployment options are sold through Cohere's sales team; this is the path for regulated-industry deployments requiring data residency.

Cloud platform availability:

• Amazon Bedrock -- Command R and Command R+ became available in Amazon Bedrock in April 2024, following initial Azure-first availability. Bedrock access uses standard AWS managed endpoints with no direct API key requirement outside of AWS IAM.
• Microsoft Azure AI Foundry -- Command R+ was available on Azure at launch, making it one of the earlier enterprise model deployments on the Azure AI platform.
• Oracle Cloud Infrastructure (OCI) -- Command models are available through OCI Generative AI; Command A and related models are the current primary offerings.

For non-commercial research, open weights are available on Hugging Face under the CC-BY-NC 4.0 license. The full-precision 104B model is at CohereLabs/c4ai-command-r-plus; a 4-bit quantized variant is at CohereLabs/c4ai-command-r-plus-4bit for smaller GPU configurations. The August 2024 refresh variant is at CohereLabs/c4ai-command-r-plus-08-2024. Community GGUF conversions are available through third-party contributors. The CC-BY-NC license prohibits commercial use; commercial deployment requires the Cohere API or an enterprise license agreement.

Comparison

Direct competitors in the enterprise text generation and RAG category, as of April 2026:

• Command A (Cohere). The immediate successor within the same lab. Command A offers 256,000-token context (double Command R+), 150% higher throughput, 23-language optimization, and comparable pricing at $2.50/$10.00 per million tokens. For new enterprise deployments, Command A is the current recommendation; Command R+ remains active for workloads already integrated against its API and for open-weights research use.
• Llama 4 (Meta AI). The leading open-weights general-purpose model as of April 2026. Llama 4 Maverick (400B total parameters, 17B active via mixture-of-experts) benchmarks higher on general-purpose evaluations and is available for self-hosted deployment under the Llama Community License. The relevant comparison for enterprise buyers is Llama 4's breadth versus Command R+'s RAG-citation specialization: Llama 4 supports fine-tuning and on-premises deployment without a per-token API cost, but RAG with structured inline citations is not a native output behavior in the way it is for Command R+.
• Mistral Large 2 (Mistral AI). A European-origin closed-weights model positioned for enterprise use with multilingual coverage and strong reasoning benchmarks. Mistral Large 2 and Command R+ compete for European enterprise customers and sovereign-AI deployments. Mistral's multilingual coverage is broadly comparable; the principal differentiation is that Cohere's RAG citation behavior is more opinionated and integrated, while Mistral Large 2 is positioned as a general-capability model that can be adapted to RAG architectures.
• GPT-5.5 (OpenAI) and Claude Opus 4.7 (Anthropic). The closed-source frontier leaders on general benchmark evaluations. Both outperform Command R+ on most composite indices, coding tasks, and complex reasoning. The comparison is relevant for enterprise buyers who are evaluating general assistant capability versus RAG-and-retrieval specialization. Command R+'s positioning does not attempt to compete on raw reasoning benchmarks with GPT-5.5 or Claude Opus 4.7; it competes on deployment flexibility, citation quality, and on-premises availability.
• DeepSeek V4 (DeepSeek). A cost-efficient open-weights Chinese-origin model that has closed the performance gap with frontier models on general benchmarks. For cost-sensitive enterprise buyers who can accept a Chinese-origin supply chain, DeepSeek V4 offers higher general benchmark scores at lower inference cost. Command R+'s differentiation against DeepSeek V4 is the RAG citation behavior, the Cohere-managed API with enterprise SLAs, and the open-weights option that does not involve the China-origin considerations that affect enterprise procurement.

Outlook

Open questions for Command R+ and the Command family over the next 6 to 18 months:

• Command A successors. Whether Cohere releases a Command B or next-generation flagship before the Aleph Alpha merger closes is unclear. Merger integration could compress the release cadence; combined engineering capacity could also accelerate it.
• Open-weights policy post-merger. The Aleph Alpha merger and Schwarz Group anchor relationship are oriented toward sovereign-AI enterprise deals. Whether the open-weights research line under Cohere Labs continues at pace or is deprioritized in favor of closed-weights enterprise products is an open question.
• Joelle Pineau's research agenda. Joelle Pineau, who joined Cohere as Chief AI Officer in August 2025 after eight years leading Meta's AI research division, oversees Cohere Labs and the research direction for the Command family. Her track record on reproducibility and responsible AI at Meta FAIR suggests a research orientation suited to RAG and agent-evaluation benchmarking.
• RAG as default infrastructure. As retrieval-augmented generation becomes a standard deployment pattern, more general-purpose models are integrating native citation and grounding behaviors. How Cohere maintains differentiation against that commoditization is the central strategic question for the Command line.
• Command R+ open-weights community. The 104B open-weights release has seeded a researcher and fine-tuner community. Whether Cohere publishes open weights for Command A or a future model at comparable scale will determine whether that community remains anchored to Cohere's family or migrates to Llama 4 or other alternatives.

Sources

• Cohere: Introducing Command R+. April 4, 2024 launch announcement by Aidan Gomez, CEO.
• Cohere Docs: Command R+ Model. Official technical specification including context window, optimized languages, and API features.
• Cohere Docs: Command A. Successor model specification and positioning relative to Command R+.
• Cohere Changelog: Command Gets Refreshed. August 2024 refresh announcement with throughput and latency improvements.
• Hugging Face: CohereLabs/c4ai-command-r-plus. Open-weights model card including parameter count, license, and capability documentation.
• AWS: Command R and Command R+ now available in Amazon Bedrock. April 2024 cloud distribution announcement.
• TechCrunch: Cohere hires Joelle Pineau as Chief AI Officer. August 2025 leadership announcement.
• VentureBeat: Cohere launches Command R+. Coverage of April 2024 launch and enterprise positioning.