# Phi-4 (14B)

> OnPrem World · https://onpremworld.com/en/models/phi-4 · Last updated: 2026-10-14 · CC BY 4.0

Microsoft's dense 14B, aimed at math and reasoning, mostly in English; its 16K context is short, but it fits a 12 GB card at 5-bit.

|  |  |
|---|---|
| Developer | Microsoft |
| Parameters | 14B |
| Native context | 16,384 tokens |
| Licence | MIT |
| Commercial use | Yes |
| KV cache | All 40 layers (standard attention) |
| Released | 2024-12 |

```
ollama run phi4:14b
```

## Memory needed

Weights + KV cache + runtime overhead, in GiB. Lower quantization = less memory, slightly lower quality.

| Quantization | 4K | 8K | 32K |
|---|---|---|---|
| Q4_K_M | 9.6 GiB | 10.4 GiB | 15.0 GiB |
| Q5_K_M | 11.0 GiB | 11.8 GiB | 16.5 GiB |
| Q6_K | 12.6 GiB | 13.3 GiB | 18.0 GiB |
| Q8_0 | 15.8 GiB | 16.6 GiB | 21.3 GiB |
| FP16 | 28.7 GiB | 29.4 GiB | 34.1 GiB |

## Which hardware runs it (8K)

| Hardware | Memory | Best quant | tok/s (est.) |
|---|---|---|---|
| GeForce RTX 5080 | 16 GB | Q6_K | 58 |
| GeForce RTX 5070 Ti | 16 GB | Q6_K | 54 |
| GeForce RTX 4090 | 24 GB | Q8_0 | 47 |
| GeForce RTX 5070 | 12 GB | Q5_K_M | 47 |
| Radeon RX 7900 XTX | 24 GB | Q8_0 | 45 |
| GeForce RTX 5090 | 32 GB | FP16 | 45 |
| RTX PRO 6000 Blackwell Workstation Edition | 96 GB | FP16 | 45 |
| GeForce RTX 4080 SUPER | 16 GB | Q6_K | 45 |
| GeForce RTX 3090 | 24 GB | Q8_0 | 44 |
| Radeon RX 9070 XT | 16 GB | Q6_K | 39 |
| Radeon RX 7900 XT | 20 GB | Q8_0 | 38 |
| RTX PRO 5000 72GB Blackwell | 72 GB | FP16 | 34 |
| Arc B580 | 12 GB | Q5_K_M | 32 |
| M5 Ultra | 512 GB | FP16 | 28 |
| GeForce RTX 5060 Ti 16GB | 16 GB | Q6_K | 27 |
| GeForce RTX 3060 (12 GB) | 12 GB | Q5_K_M | 25 |
| RTX 6000 Ada Generation | 48 GB | FP16 | 24 |
| Arc Pro B60 | 24 GB | Q8_0 | 21 |
| M5 Max (18-core CPU / 32-core GPU) | 36 GB | Q8_0 | 20 |
| Radeon RX 9060 XT 16GB | 16 GB | Q6_K | 19 |
| RTX A6000 | 48 GB | FP16 | 19 |
| M3 Ultra | 512 GB | FP16 | 19 |
| M4 Max (14-core CPU / 32-core GPU) | 36 GB | Q8_0 | 18 |
| GeForce RTX 4060 Ti 16GB | 16 GB | Q6_K | 17 |
| Radeon AI PRO R9700 | 32 GB | FP16 | 16 |
| Arc Pro B70 | 32 GB | FP16 | 15 |
| M5 Max (18-core CPU / 40-core GPU) | 128 GB | FP16 | 14 |
| M4 Max (16-core CPU / 40-core GPU) | 128 GB | FP16 | 13 |
| M6 (16GB configuration, 153 GB/s) | 16 GB | Q5_K_M | 10.0 |
| M6 (24GB/32GB configurations, 170 GB/s) | 32 GB | Q8_0 | 7.4 |
| M5 Pro | 64 GB | FP16 | 7.1 |
| M5 | 32 GB | Q8_0 | 6.7 |
| RTX Spark (N1X) laptops & compact desktops | 128 GB | FP16 | 6.4 |
| M4 Pro | 64 GB | FP16 | 6.3 |
| DGX Spark | 128 GB | FP16 | 5.9 |
| Ryzen AI Max+ PRO 495 (Gorgon Halo) systems | 192 GB | FP16 | 5.9 |
| Ryzen AI Max+ 395 (Strix Halo) systems | 128 GB | FP16 | 5.5 |

## Frequently asked questions

### How much VRAM does Phi-4 (14B) need?

With an 8K context, about 10.4 GiB at Q4_K_M and 16.6 GiB at Q8_0 (weights + KV cache + runtime overhead). Longer contexts need more memory for the KV cache.

### Can Phi-4 (14B) run on an GeForce RTX 4090?

Yes. Phi-4 (14B) fits on the GeForce RTX 4090 (24 GB) at Q8_0 with an 8K context, using about 16.6 GiB, at an estimated 47 tokens/s.

### Can I run Phi-4 (14B) on a Mac?

Yes. Phi-4 (14B) fits on the M3 Ultra (96 GB) at FP16 with an 8K context, using about 29.4 GiB, at an estimated 19 tokens/s.

### Can Phi-4 (14B) be used commercially?

Phi-4 (14B) is released under MIT. The licence allows commercial use (read its conditions).

## Sources

- https://huggingface.co/microsoft/phi-4
- https://huggingface.co/microsoft/phi-4/resolve/main/LICENSE
- https://huggingface.co/microsoft/phi-4/resolve/main/config.json
- https://huggingface.co/api/models/microsoft/phi-4?expand[]=safetensors
- https://ollama.com/library/phi4/tags