Command R+
Open weightsby Cohere·Other·Released
Cohere's RAG-first enterprise flagship — strong citations, on-prem, BYOC deployment.
About this model
Command R+ (August 2024) is Cohere's flagship — purpose-built for enterprise RAG and tool-use workflows. Cohere has positioned itself explicitly as 'the enterprise AI company,' and Command R+ reflects that focus: strong citation accuracy, on-prem and bring-your-own-cloud deployment options, SOC 2 and HIPAA certifications, and a tool-use API designed for production agent workflows.
On general benchmarks Command R+ trails the top US frontier models, but for the enterprise RAG use case it was designed for — where citation accuracy and grounding matter more than raw reasoning — it's often the preferred choice.
Notably, Cohere releases weights under a CC-BY-NC license for research, which is more open than the major closed-weights frontier labs.
Strengths
- •Best-in-class RAG grounding and citation accuracy
- •On-prem and BYOC deployment supported out of the box
- •SOC 2, HIPAA, with FedRAMP work in progress
- •Strong multilingual: 23 languages with explicit performance parity
- •Research weights available under CC-BY-NC
Limitations
- •Trails GPT-4o / Claude 3.5 Sonnet on general benchmarks
- •Smaller developer ecosystem than OpenAI / Anthropic
- •Weights are research-only (no commercial use without API)
When to use it
- →Regulated-industry enterprise RAG (finance, healthcare, government)
- →On-prem deployments under strict data-residency rules
- →Customer-facing assistants needing precise citations
- →BYOC deployments on customer AWS / Azure / GCP
Architecture & training
104B-parameter dense transformer. Cohere has chosen dense over MoE for serving predictability — enterprise customers prefer consistent per-request latency over the variable behaviour MoE models can exhibit under load. Pretraining data is heavily curated for citation-friendly content (academic papers, technical documentation, structured knowledge bases).
Benchmarks
| Benchmark | Score | Bar |
|---|---|---|
| MMLU | 75.7 | |
| HumanEval | 70.1 |