Microsoft: Phi 4

Phi-4 is a dense decoder-only Transformer model designed to push the boundaries of what small language models can achieve in complex reasoning. With 14 billion parameters, it is engineered to excel in logic-heavy domains such as mathematics and advanced language processing. The design intent focuses on creating a high-performance building block for generative AI applications, particularly in environments where memory, compute, or latency constraints are primary concerns. By prioritizing high-quality data over sheer scale, the model provides a robust solution for developers seeking sophisticated reasoning logic in a compact, efficient footprint.

The model's performance is rooted in a rigorous training curriculum that utilizes a blend of synthetic datasets, filtered public domain content, and academic materials. To refine its output, the model underwent an extensive post-training phase incorporating supervised fine-tuning and direct preference optimization, which ensures precise instruction adherence and safety. These advancements allow it to outperform significantly larger models on various benchmarks, demonstrating a superior size-to-quality ratio. As part of a broader family of models, Phi-4 serves as a foundation for future research and practical deployment, offering a reliable tool for developers building AI-powered features that require deep analytical capabilities.