phi-4
by microsoft
14.7B params · text-generation · 2.2k likes · 959.0k downloads
phi-4 is a 14.7B parameter model. At Q4 quantization it requires 7GB of VRAM. It runs comfortably on GeForce RTX 4090 (89 tok/s), GeForce RTX 5090 (134 tok/s), M4 Max 128GB (32 tok/s).
GPU compatibility
| GPU | VRAM | Q4 Decode | Verdict |
|---|---|---|---|
| GeForce RTX 4090 | 24GB | 89 tok/s | comfortable |
| GeForce RTX 5090 | 32GB | 134 tok/s | comfortable |
| M4 Max 128GB | 128GB | 32 tok/s | comfortable |
| M4 Pro 48GB | 48GB | 16 tok/s | tight |
| M4 Pro 24GB | 24GB | 16 tok/s | tight |
| A100 PCIe 80 GB | 80GB | 164 tok/s | comfortable |
| H100 SXM5 80 GB | 80GB | 330 tok/s | comfortable |
| GeForce RTX 3090 | 24GB | 79 tok/s | comfortable |
| Radeon RX 7900 XTX | 24GB | 65 tok/s | comfortable |
| GeForce RTX 4080 | 16GB | 63 tok/s | comfortable |