Phi-4

Phi-4 is a 14-billion parameter dense decoder-only Transformer designed to deliver strong reasoning and instruction-following capabilities within a compact footprint. Built by Microsoft Research, the model was trained on a curated blend of synthetic data, filtered public domain content, and academic books and Q&A datasets totaling 9.8 trillion tokens—a deliberate strategy to prioritize quality over quantity in the training data. This approach reflects the broader Phi family philosophy of extracting maximum capability from smaller, higher-quality datasets rather than scaling blindly.

The model underwent a rigorous post-training pipeline combining supervised fine-tuning and direct preference optimization to sharpen instruction adherence and safety behavior. Trained across approximately 1,920 H100-80G GPUs over 21 days, Phi-4 lands under an MIT license, making it accessible for both research exploration and commercial deployment. Its sweet spot lies in latency-sensitive and memory-constrained environments where efficient inference matters, with particular strength in reasoning-intensive tasks and English-language applications. As an open-weights model, it serves as a flexible building block for developers looking to integrate capable language understanding into custom AI products without relying on fully proprietary infrastructure.