Llama 3.1 405B Instruct, Mistral Large 2, and Qwen 3 72B Instruct span a range of parameter counts and publisher strategies, but all target high-quality open-weights inference for production use. Meta's 405B (July 2024, Llama 3 community license) sits at the frontier of what's accessible in open weights; Mistral Large 2 (July 2024, Mistral Research license) and Qwen 3 72B (Alibaba, 2025, Qwen license) compete at the 70B-class tier with different multilingual and licensing profiles. Qwen 3 72B is the most current of the three, bringing 2025-generation instruction tuning and strong multilingual performance across CJK and Arabic alongside 131K context. For teams serving non-English users, its multilingual advantage over both alternatives is tangible. The Qwen commercial license permits deployment without restrictions. Mistral Large 2 delivers competitive multilingual performance across European languages and solid coding evals at the 70B scale. The Mistral Research license limits fully commercial use without an enterprise agreement — a real constraint for teams that need unrestricted deployment or fine-tuning rights. For organizations in an existing Mistral relationship, the quality-per-token profile remains competitive with the Qwen 3 72B. Llama 3.1 405B is in a different operating tier. At 405 billion parameters it handles tasks that genuinely expose 70B limitations — complex multi-step reasoning, long-form synthesis, advanced coding workflows. Multi-GPU infrastructure requirements and thinner provider availability make it a targeted choice rather than a volume inference default. Per-token cost is substantially higher than either 70B-class alternative. Pick Qwen 3 72B for general-purpose or multilingual production deployments. Pick Mistral Large 2 within an existing Mistral commercial agreement. Pick Llama 3.1 405B only when task complexity demonstrably saturates 70B-class models and the infrastructure cost is justified.