Overview

Phi-4 continues Microsoft’s Phi series of small language models, which aim to show that careful data curation can rival much larger models. Its training relied heavily on synthetic data designed to teach reasoning, and Microsoft reported strong results on math competition problems relative to its size.

Highlights

  • 14 billion parameters
  • Trained with a large share of curated synthetic data
  • Strong math and reasoning performance for its size
  • MIT license permitting commercial use
  • Runs on a single GPU or, quantized, on high-end laptops

Typical uses

On-device and low-cost deployments, reasoning-heavy tasks with tight budgets, and research on data-efficient training.

Things to consider

  • Shorter context window than many larger models.
  • Primarily English; weaker on factual knowledge than larger models.
  • Smaller variants and reasoning versions were released later; compare them.

Specifications are taken from the provider’s official announcement, documentation or model card. Benchmark results depend on evaluation settings; see how AI benchmarks work.