Three models sharing a 131K context window but representing a wide spectrum of parameter counts and capability profiles. Meta's Llama 3.1 70B Instruct, released July 2024, was a milestone: the first 70B-class open model with a genuine 131K context window. MMLU scores around 79–80 and solid general-instruction-following made it the default open-weights workhorse through late 2024. Provider coverage is among the broadest for any open model. The practical note for 2026: Llama 3.3 70B from December 2024 improves instruction-following at the same footprint, and new 70B deployments should generally prefer 3.3 unless pinning to a specific weight hash. Phi-3 Medium 128K at 14B parameters is the cost-efficiency outlier. Microsoft's synthetic-data training approach closes a surprising portion of the MMLU gap to 70B models on reasoning-heavy tasks, making it attractive when per-token cost is tightly constrained. It lags on multilingual evals, open-ended generation, and knowledge breadth — but for extraction, summarization, or structured QA, the per-parameter efficiency is real. MIT license. Qwen 2.5 72B Instruct is the September 2024 Alibaba flagship, since superseded by Qwen 3 72B in April 2025. It remains widely deployed because fine-tuned adapters and cached prompt distributions rarely migrate quickly. MMLU and multilingual scores are competitive against Llama 3.1 70B, and CJK coverage exceeds either peer in this comparison. Pricing has started to soften as providers shift capacity toward Qwen 3. The Qwen commercial license applies. Pick Llama 3.1 70B (or preferably 3.3 70B) for the broadest hosted coverage and reliable general-purpose performance. Pick Qwen 2.5 72B when multilingual workloads, existing adapters, or pinned checkpoints tie you to the 2.5 generation. Pick Phi-3 Medium 128K when per-token cost is the primary constraint on a reasoning-focused task.