Phi-4

Phi-4 is the fourth iteration of Microsoft's small-model research line. It's a 14B dense model trained with a heavy emphasis on synthetic data — about 40% of the training corpus is model-generated, carefully curated for educational and reasoning content.

The synthetic-data thesis

The Phi line argues that data quality dominates data quantity. Phi-4 was trained on roughly 10 trillion tokens, far less than its 14B-class peers, but performs disproportionately well on math, science, and reasoning benchmarks. On GPQA and MATH it competes with much larger models.

What it's good at

STEM reasoning, well-structured single-turn responses, code (especially algorithmic), and producing clean explanations. It's a popular choice as a tutor-style assistant or as a base for fine-tuning on technical content.

What to watch for

16K context is short. The model is also more "academic" in tone — it tends to be slightly stiffer in casual conversation than its peers. Multilingual performance is limited; treat it as English-first.

License

MIT — fully permissive.