OLMo 2 13B Instruct, Phi-3 Medium 128K, and Qwen 3 14B Instruct are three 13–14B class models from very different organizational contexts: an academic AI lab, a major cloud vendor, and a hyperscaler. What differentiates them is not primarily parameter count but training philosophy, licensing, and the use cases each organization optimized for. OLMo 2 13B Instruct from Allen AI was released November 2024 with all training data, weights, and training code published under Apache 2.0. Full-stack openness — the Dolma corpus, OLMo-mix training framework, and public evaluation results — makes it the most reproducible model in the sub-20B class. Academic teams and ML researchers who need to audit or replicate training results consistently reach for it. The practical constraint is severe: the 4K context window rules out most RAG architectures, and hosted coverage is thin. This is a research and fine-tuning base, not a general production endpoint. Microsoft's Phi-3 Medium 128K is a 14-billion-parameter model released May 2024 under MIT license, trained on heavily filtered textbook-quality synthetic data. MMLU scores and GSM8K accuracy exceed most 14B peers, narrowing the gap to 70B-class models on reasoning-heavy tasks. The 131K context window handles long-document summarization and extended code reviews without chunking. Hosted coverage skews toward Azure AI. Qwen 3 14B Instruct from Alibaba has 131K context, strong CJK multilingual capability, and latency and throughput comparable to Llama 3.1 8B at the 14B scale. For teams with East Asian user traffic or multilingual document workloads, it is the practical first choice in this size bracket. The Qwen license permits commercial use with attribution. Pick OLMo 2 13B for research, reproducibility, and fine-tuning where full Apache 2.0 data provenance matters. Pick Phi-3 Medium 128K for reasoning-intensive workloads on Azure. Pick Qwen 3 14B for production deployments requiring multilingual support and long-context capability.