Phi

Trained hard on high-quality material rather than raw volume, so the small models punch above their size on reasoning. MIT licensed.

Small models have a ceiling. Excellent benchmarks do not always survive contact with a real task, especially anything requiring world knowledge.