Three large open-weights models — but with very different production deployment profiles despite all being in the 72B-to-405B parameter range. Llama 3.1 405B Instruct is 405 billion dense parameters, 131K context, and Meta's broadest open-weights model at its July 2024 release. MMLU scores rank it near the top of open models. Provider coverage is limited to those who can provision multi-GPU infrastructure, but commercially viable hosting does exist on Lambda Labs, Fireworks, and others. The Llama 3 community license supports commercial use with attribution. Nemotron-4 340B Instruct from NVIDIA is a 340B dense model released June 2024, positioned explicitly as a synthetic-data-generation engine rather than a general-purpose deployment target. At 340B parameters in a non-MoE architecture it is rare at this scale in 2026, but the 4K context ceiling is a hard constraint that disqualifies it from long-document or extended-conversation workloads. Hosting is concentrated on NVIDIA's NIM service, limiting provider flexibility. The NVIDIA Open Model License is not OSI-approved — review commercial terms with legal. If you are generating synthetic fine-tuning datasets at scale and need a large dense reference model, this is the use case. For anything else, the context ceiling kills it. Qwen 3 72B Instruct, released April 2025, covers the practical workload space that Nemotron's context ceiling cuts off and at significantly lower serving cost than 405B. Strong MMLU, HumanEval, and multilingual scores under the Qwen commercial license. Pick Llama 3.1 405B for frontier open-weights capability with flexible commercial licensing and genuine long-context support. Pick Nemotron-4 340B specifically for synthetic dataset generation tasks where its training was targeted. Pick Qwen 3 72B for everyday high-performance inference at a fraction of the cost.