Phi
A 14B dense decoder trained for reasoning.
Phi-4 is a dense decoder trained for reasoning rather than scaled for it — the argument being that a well-trained 14B model can answer questions usually put to much larger ones. Whether that holds for your workload is a question for your evaluation, not for this page; what this page can tell you is that it appears in the v0.5.1 matrix with the function and performance evidence reported by this release snapshot.
Phi-4 14B
- Kind: Language model
- Task: Text generation, reasoning
- Parameters: 14B
- Architecture: Dense transformer
- Quantization: int4 RTN, block 32
- Base model: microsoft/phi-4
- Validation status: Validated
- Measured in v0.5.1: 24.01 tokens/s, 2.52 s to first token, at a 2048-token prompt — all three lengths