An enterprise RAG specialist, a frontier-scale dense model, and a balanced multilingual 72B — each makes sense for a different primary workload. Cohere's Command R+ is a 104B-parameter model released April 2024, built explicitly for retrieval-augmented generation and multi-step tool use. It scores well on grounding and function-calling benchmarks, which reflects deliberate architectural and fine-tuning investment in those capabilities rather than simply scaling up. The 131K context window handles long retrieval contexts. The significant constraint is licensing: Command R+ uses Cohere's CC-BY-NC license, which prohibits commercial deployment through third-party inference hosts. Production use channels through Cohere's own API. Teams evaluating Command R+ should factor in the vendor dependency alongside the RAG-quality argument. Llama 3.1 405B Instruct is the largest dense open-weights model in this group at 405B parameters, 131K context, and Llama 3 community license terms that allow commercial use with attribution. MMLU scores near the top of the open-weights tier at launch, broad provider support, and the ability to self-host on sufficient hardware. The cost of inference is significantly higher than either peer — justified when the task genuinely requires frontier-level capability. Qwen 3 72B Instruct from April 2025 covers MMLU, code, and multilingual benchmarks comprehensively at a fraction of 405B pricing. For most RAG pipelines that need broad knowledge retrieval rather than Cohere's specialized grounding stack, Qwen 3 72B is the more practical default. Pick Command R+ when you need Cohere's grounding stack and are willing to route through their API. Pick Llama 3.1 405B when frontier capability and licensing flexibility to self-host are the priority. Pick Qwen 3 72B for general-purpose high-performance inference at reasonable 72B pricing.