This is a comparison across three different scale philosophies: massive dense weights, efficient small-model training data, and a balanced mid-tier with broad multilingual coverage. Llama 3.1 405B Instruct is Meta's largest openly licensed model from July 2024 — 405 billion dense parameters, 131K context, and the Llama 3 community license. It scores near the top of open-weights models on MMLU and complex reasoning evaluations. The cost to run 405B is substantial: hosted pricing is significantly higher than 70B-class alternatives, and multi-GPU requirements limit which providers carry it. It makes the most sense for workloads where raw capability matters more than cost-per-token — synthesis, complex document analysis, or agentic tasks that genuinely benefit from scale. Phi-3 Medium 128K brings 14 billion parameters trained on Microsoft's curated synthetic corpus, delivering MMLU and GSM8K scores that approach several 70B models on reasoning-heavy benchmarks. The 131K context window is present, and MIT licensing removes commercial friction. The tradeoff is coverage: it underperforms on open-ended generation, creative tasks, and anything requiring broad world knowledge over reasoning depth. Provider availability is more limited than either peer. Qwen 3 72B Instruct, released April 2025, occupies the middle ground well. At 72B parameters with a 131K context window, it covers MMLU, multilingual, and code benchmarks with fewer gaps than Phi-3 Medium while costing a fraction of 405B inference. The Qwen license supports commercial deployment. Pick Llama 3.1 405B when your task genuinely requires frontier-level open-weights capability. Pick Qwen 3 72B for strong all-around performance at reasonable 72B pricing. Pick Phi-3 Medium 128K when per-token cost is the constraint and the task is structured reasoning or QA.