Llama 3.1 405B Instruct, Llama 3.1 70B Instruct, and Llama 3.3 70B Instruct are all Meta open-weights models released under the Llama 3 community license, all with 131K context windows. The 405B and 70B variants launched together in July 2024; Llama 3.3 70B arrived in December 2024 as an improved instruct-tuned version at the 70B footprint. The practical question here is whether the 405B capability ceiling is worth the hosting premium over the updated 70B. Llama 3.3 70B is the cleaner baseline at the 70B tier. The December 2024 instruction tuning improvements deliver measurably better multi-turn coherence, tool-use adherence, and structured-output reliability compared to the 3.1 70B, at roughly equivalent inference cost. For most production workloads, Llama 3.3 70B should be the default comparison point, not the older 3.1 variant. Llama 3.1 70B still runs at comparable cost and hardware requirements, but its instruction-following quality has been overtaken by its successor. Teams should be running 3.3 for any new deployment unless pinned to a specific weight hash. Llama 3.1 405B sits at the upper end of what any open-weights provider offers at scale. At 405 billion parameters it handles complex reasoning chains, extended code generation, and long-document analysis tasks that reveal quality gaps in 70B models. Multi-GPU inference requirements mean hosted per-token pricing is meaningfully higher and provider availability thinner. It is not the right pick for general-purpose volume inference. Pick Llama 3.3 70B for new general-purpose deployments at the 70B tier. Use Llama 3.1 70B only when checkpoint reproducibility is required. Choose Llama 3.1 405B when task complexity visibly saturates 70B performance and you can justify the multi-GPU infrastructure cost.