Three competitive open-weights flagships for teams that want quality without paying proprietary API rates. Llama 3.3 70B Instruct is Meta's December 2024 70B dense model, 131K context, and the most permissive licensing of the three under the Llama 3 community license. It is the default choice when provider flexibility and uncomplicated deployment matter — dozens of hosts carry it, and the instruction-following quality improvement over 3.1 70B is real. Fine-tuning on this base is practical on two A100s, which matters for teams customizing toward domain-specific tasks. Mistral Large 2 is Mistral AI's 123B-parameter model from July 2024, with a 128K context window and genuine investment in European multilingual quality, function calling, and structured JSON output. The 53B parameter gap over Llama 3.3 70B translates into a measurable quality advantage on complex multi-step tasks and multilingual European text. The Mistral Research License complicates self-hosting at scale — most production deployments run through Mistral's own API, which adds a dependency but provides reliability and tooling. Qwen 3 72B Instruct is Alibaba's April 2025 open model, matching Mistral Large 2 on English benchmarks and surpassing both competitors on CJK and Arabic language tasks. The 131K context window is comparable. The Qwen commercial license permits self-hosting and API use. For global products with significant non-English user traffic, it is frequently the best fit without a per-token premium. Pick Llama 3.3 70B for Apache-near permissiveness and widest provider coverage. Pick Mistral Large 2 when European language quality and Mistral's function-calling API ecosystem are worth the managed-API dependency. Pick Qwen 3 72B when multilingual Asian and Middle Eastern languages are a core product requirement.