Phi-4

Permissive

Microsoft · Phi · Released 2024-12

14BText

A data-quality-first small model that punches far above its 14B size on math and reasoning benchmarks, fully MIT licensed.

Strengths

  • +MIT license, unrestricted use
  • +Strong math/reasoning benchmark scores relative to size
  • +Quantizes down to a single consumer GPU

Limitations

  • -16K context window is small versus 2024-25 peers
  • -Weaker general world-knowledge breadth than larger models

License

MIT

use commercially with attribution niceties

Hardware

runs on a good consumer GPU (or Apple Silicon) with quantization

Links

Stats

microsoft/phi-4— downloads·— likes

via Hugging Face

Related models