Phi-4
phi-4
Microsoft · open-permissive mit
| Parameters (B) | 14.66 |
|---|---|
| Architecture | dense |
| Layers / KV heads / head dim | 40 / 10 / 128 |
| Context window | 16384 |
| Modalities | in: text; out: text |
| GGUF | available |
| Commercial use | yes |
Minimum hardware (estimate)
At Q4_K_M and 8k context, total working set ≈ 11.1 GiB. On a machine with 64 GiB system RAM and no discrete GPU: CPU-only, slow. Limiting factor: No GPU VRAM; fits in system RAM but inference will be CPU-bound.
Scores
| Benchmark | Value | Uncertainty | Run by | Source |
|---|---|---|---|---|
| GPQA Diamond | 56.1% | unknown | third-party | epoch-ai |
| MATH Level 5 | 64.9% | unknown | third-party | epoch-ai |
| OTIS Mock AIME 2024-2025 | 13.8% | unknown | third-party | epoch-ai |
Provenance
- identity, license, architecture, parameter count: https://huggingface.co/microsoft/phi-4 (retrieved 2026-08-10)
HF: microsoft/phi-4 · snapshot 2026-08-10