Command R+

What is Command R+?

Command R+ is Cohere’s flagship large language model, released in April 2024 as the most capable model in the Command R family. Unlike general-purpose models optimized primarily for chat, Command R+ was architected specifically for retrieval-augmented generation (RAG) and enterprise tool use — the two most demanding production AI workloads in corporate environments. With 104 billion parameters and a 128,000-token context window, it sits in the same capability class as GPT-4-level models while being purpose-built for the structured, grounded information retrieval patterns that enterprise applications require.

What distinguishes Command R+ from competitors is its native training for multi-hop RAG — the ability to query multiple retrieved documents, reconcile conflicting information, synthesize a coherent answer, and provide inline citations that cite which passage each fact came from. Most models treat RAG as an afterthought, injecting retrieved context into a general-purpose prompt template. Command R+ was explicitly trained on RAG-style data, giving it substantially better grounding accuracy, fewer hallucinations on retrieved content, and more reliable citation formatting than models not purpose-trained for this workflow.

Command R+ also features strong multilingual capabilities across 10+ languages — English, French, Spanish, Italian, German, Portuguese, Japanese, Korean, Arabic, and Chinese — making it a preferred choice for multinational enterprise deployments that need a single model to operate consistently across regional teams without language-specific fine-tuning.

Key Capabilities

• Multi-hop RAG with inline citations: Natively trained to retrieve from multiple sources, synthesize answers, and emit structured citations linking each claim back to the specific retrieved document chunk — dramatically reducing hallucination in knowledge-intensive tasks.
• Tool use and function calling: Reliable structured output for multi-step agentic workflows where the model must decide which tools to invoke, in what order, and how to interpret their outputs to complete a goal.
• 128k token context window: Handles large enterprise documents — lengthy contracts, technical specifications, financial filings — in a single context without chunking or truncation artifacts.
• Multilingual performance: Maintains strong performance across 10+ languages on RAG-style tasks, avoiding the multilingual performance degradation common in models primarily trained on English data.
• Low hallucination rate on grounded tasks: Cohere’s internal evals show Command R+ produces fewer fabricated citations and fewer off-document claims than competing models of similar size when operating in a RAG setup.
• Cohere platform integration: Native first-class support in Cohere’s Embed + Rerank + Command pipeline, allowing teams to build fully integrated RAG stacks on a single vendor with consistent API patterns.

Ideal Use Cases

• Enterprise knowledge management: Powering internal Q&A systems over large corporate document repositories — policy wikis, legal archives, product documentation — where accurate citation of sources is non-negotiable.
• Legal and compliance research: Automating discovery and contract review workflows where the model must search across multiple documents and produce cited, traceable summaries rather than free-form answers.
• Customer support with knowledge bases: Grounding support agent responses in product manuals, FAQ databases, and ticketing history to eliminate hallucinated instructions that could harm customer relationships.
• Multilingual enterprise deployments: Providing consistent AI assistant capabilities across regional offices in different countries without the overhead of maintaining separate language-specific models.
• Agentic research pipelines: Acting as the reasoning core of multi-step agent workflows that must execute searches, process results, make decisions, and produce structured reports with traceable evidence chains.

Limitations & Caveats

• Coding performance: Command R+ scores 61% on HumanEval, trailing significantly behind top coding-optimized models like Claude 3.5 Sonnet (92%) or DeepSeek Coder V2 (90.2%). It is not the right choice for complex software development tasks.
• License restrictions: The C-UDA (Cohere Universal Developer Agreement) license permits commercial use but with specific restrictions around certain high-risk applications. Enterprise deployments should review the terms carefully, particularly for regulated industries.
• Weaker general reasoning vs. frontier models: While Command R+ excels at RAG and grounded tasks, it lags behind GPT-4o and Claude 3.5 Sonnet on open-ended reasoning, creative tasks, and mathematical problem-solving outside a retrieval context.
• Context window effectiveness: Like all transformer models, performance on information in the middle of very long contexts (the “lost in the middle” phenomenon) can degrade for very long 128k inputs, particularly when critical information is surrounded by large amounts of irrelevant retrieved text.

Architecture Notes

Command R+ uses a standard decoder-only transformer architecture with grouped-query attention (GQA) for efficient inference at 104B scale. The model is optimized for structured input formats: it was fine-tuned with special tokens and training data designed specifically for the SYSTEM:, USER:, and CHATBOT: prompt roles, with additional special tokens for tool-use schemas and RAG citation formats. This explicit structural training is what enables its superior citation consistency compared to models that rely on prompt engineering alone for structured outputs.

Cohere trains Command R+ with a multi-objective approach: standard next-token prediction, a citation grounding objective that rewards sourcing answers from provided documents, and a tool-calling alignment stage that teaches the model to emit function call schemas in structured JSON. The model is available via Cohere’s managed API and through major cloud providers including AWS Bedrock and Azure AI Studio, with both full-precision and optimized inference options. Cohere also provides the model through their on-premise deployment offering for enterprises with data residency requirements.