Three 128K-context models at different parameter counts covering the practical mid-tier of open-weights inference. The spread — 14B, 70B, and 72B — matters more than the numbers suggest because training data quality and architecture choices cut across raw scale. Llama 3.3 70B Instruct is Meta's December 2024 70B refresh, targeting better instruction-following than Llama 3.1 70B at the same footprint. It holds the familiar 131K context and Llama 3 community license, and is already the recommended replacement over 3.1 70B for new deployments. Provider coverage is among the widest of any open model, giving you genuine flexibility on cost and latency. On standard evals, it closes a portion of the gap to 405B-class models that existed in 3.1. Phi-3 Medium 128K, at 14 billion parameters, is the outlier on scale in this group. Microsoft's training data approach produces MMLU and GSM8K scores that match some 70B competitors on reasoning tasks, while costing substantially less per token. The gap shows in open-ended generation quality and broad knowledge coverage — GPQA-style science reasoning and MT-Bench conversational quality both reflect the smaller parameter budget. MIT license, Azure AI primary hosting. Qwen 3 72B Instruct is Alibaba's April 2025 flagship 72B, covering multilingual, code, and reasoning benchmarks more evenly than Phi-3 Medium while offering competitive pricing against Llama 3.3 70B across providers like Together AI, Fireworks, and Groq. Pick Llama 3.3 70B for the broadest provider choice and reliable instruction-following at 70B scale. Pick Qwen 3 72B when multilingual coverage or code tasks are part of the workload. Pick Phi-3 Medium 128K when per-token cost is a hard constraint and the task is structured reasoning.