Open-weight family · Fast-moving · Intermediate

Phi (Microsoft)

Microsoft Research's small models, built to show that carefully curated training data can outperform brute scale.

What Phi (Microsoft) is

The Phi line emphasises textbook-quality and synthetic curricula, achieving reasoning performance well above what their parameter counts would suggest.

How it works

Weights are published for local use and fine-tuning, with variants covering vision and longer context.

Why it matters

Phi is the standard reference for the argument that data quality beats parameter count at small scale.

Common uses

→On-device reasoning
→Cost-sensitive production tasks
→Research on data curation

Strengths

✓Excellent quality per parameter
✓Cheap to run

Watch for

✓Narrower world knowledge
✓Weaker on long open-ended tasks

Continue exploring

Microsoft Small Language Model Dataset Curation Synthetic Data