Three models that sell to different buyers. Command R+ is Cohere's 104B-parameter enterprise model, released April 2024 with a 131K context window and explicit design investment in retrieval-augmented generation and multi-step tool use. Cohere built proprietary RAG evaluation suites into the development process, and the function-calling benchmark performance reflects that investment. The catch: the CC-BY-NC license prohibits commercial deployment through third-party hosts, which means production workloads run through Cohere's own API. If you are buying the RAG tooling and the managed enterprise service, the model price makes more sense. If you want a self-hostable alternative, the license rules it out. DeepSeek V3.2 is a 671B MoE from May 2025 with roughly 37B active parameters per forward pass and a 128K context window. Its benchmark numbers on code, math, and general reasoning exceed Command R+ across most evaluations. DeepSeek's own commercial license is permissive for most use cases, and per-token costs across providers like DeepInfra and Fireworks are significantly lower than Command R+ via Cohere's API. Llama 3.3 70B Instruct is the most permissive option — Meta's December 2024 dense 70B with 131K context and the Llama 3 community license. It lacks Command R+'s dedicated RAG investment but works well with standard retrieval pipelines and runs cheaply across dozens of hosted providers. Pick Command R+ when you need Cohere's enterprise RAG ecosystem and are willing to work within their managed API. Pick DeepSeek V3.2 when quality ceiling and cost-efficiency are paramount and you can manage the licensing review. Pick Llama 3.3 70B for flexible, permissively licensed general deployment with broad provider choice.