Command R+ (08-2024)

Command R+ is Cohere's flagship open-weight release: a 104-billion-parameter dense model purpose-built for retrieval-augmented generation, tool use, and grounded enterprise workflows. The 08-2024 refresh improved instruction-following and refusal handling over the original April 2024 release.

What sets it apart

Most open-weight models treat RAG as an afterthought. Command R+ was trained with explicit document-citation tokens and a tool-call schema as first-class outputs. When given a set of documents, it produces grounded answers with inline citations to specific passages — without needing prompt engineering hacks.

What it's good at

Enterprise document QA, multi-step tool use, multilingual workflows (10 languages with strong coverage), and any task where you need the model to cite its sources. Performance on creative tasks is solid but not its strong suit.

Running it locally

Full FP16 weights are ~208 GB. Q4_K_M GGUF is ~63 GB. Practical for a multi-GPU workstation but not for a single consumer card; most users access it through Cohere's hosted API or a self-hosted vLLM deployment.

License

CC-BY-NC 4.0 — non-commercial only. For commercial use you must obtain a separate license from Cohere. This is the most restrictive license in the catalog; if commercial use matters, look at Llama 3.3 or Qwen2.5 instead.