This comparison spans three genuinely different architectures at three price points, making it one of the more practically useful comparisons in the sub-frontier tier. DeepSeek V3.2 is a mixture-of-experts model from DeepSeek, the May 2025 successor to V3, with a reported ~30% reduction in inference pricing over its predecessor. It routes each token through a subset of a large total parameter count, making its effective inference cost much lower than a comparably capable dense model. On coding, math, and general reasoning benchmarks it punches well above what its inference cost implies, which is why it drew significant attention at launch. The 131K context window is accessible across DeepInfra, Fireworks, and OpenRouter. Verify DeepSeek's commercial license terms before deploying. Phi-3 Medium 128K, at 14 billion dense parameters, trades breadth for efficiency. Microsoft's synthetic training data drives MMLU and GSM8K scores above most 14B peers, but the model does not cover the full range of benchmarks where DeepSeek V3.2 and Qwen 3 72B compete. Its primary advantage is cost: at 14B with MIT licensing, it is the lowest-cost option in this comparison for structured reasoning tasks. Provider availability skews toward Azure AI. Qwen 3 72B Instruct from Alibaba covers the broadest benchmark surface in this group — strong multilingual, code, and reasoning coverage — at 72B parameters with a 131K context window and Qwen commercial license. It is the most predictable all-rounder of the three. Pick DeepSeek V3.2 when raw benchmark performance per inference dollar is the primary metric and you can tolerate MoE operational complexity. Pick Qwen 3 72B for reliable breadth across multilingual and coding workloads. Pick Phi-3 Medium 128K when the task is narrow, structured, and cost-sensitive.