Hardware guide · October 2026
What GPU do you need to run Phi-4?
Phi-4 is a 14.7B dense model. Here is exactly how much memory it needs and which GPUs and Macs run it well.
The short answer
To run Phi-4 at Q4_K_M with 8K tokens of context you need about 12GB of VRAM (9.1GB of weights plus KV cache and overhead). The cheapest way in is a 12GB card like the RTX 3060 12GB (~$270 used). On a Mac or other unified-memory system, the Ryzen AI Max+ 395 runs it at ~12 tokens/sec.
- VRAM at Q4_K_M
- ~12GB
- GPUs that run it
- 35
- Parameters
- 14.7B
- CPU only (est.)
- ~4.9 tok/s
Our recommendations
The best GPUs for Phi-4
Minimum
RTX 3060 12GB12GB · ~$270 used
- Quant
- Q3_K_M
- Speed (est.)
- ~34 tok/s
The smallest card that runs it at Q3_K_M with 8K context.
Fastest
RTX 509032GB · ~$4,300
- Quant
- Q8_0
- Speed (est.)
- ~84 tok/s
The fastest single GPU in our database for this model.
Mac / unified memory
Ryzen AI Max+ 395128GB unified · ~$3,300
- Quant
- Q8_0
- Speed (est.)
- ~12 tok/s
Lots of memory per dollar and silent, but slower than a GPU with the same capacity.
We may earn a commission when you buy through links on this page. It never changes what we recommend. How we make money
Memory
Phi-4 VRAM requirements
Memory needed at each quantization level. Q4_K_M is the best balance of size and quality for most people; move up to Q5 or Q6 if you have the room.
| Quantization | Quality | Weights | Total @ 8K ctx |
|---|---|---|---|
| Q2_K | Rough | 5.6 GB | 8.0 GB |
| Q3_K_M | Usable | 7.3 GB | 9.8 GB |
| Q4_K_M | RecommendedDefault | 9.1 GB | 12 GB |
| Q5_K_M | Very good | 11 GB | 13 GB |
| Q6_K | Near-lossless | 12 GB | 15 GB |
| Q8_0 | Practically lossless | 16 GB | 18 GB |
| BF16 | Full precision | 29 GB | 32 GB |
Context
How context length changes the memory you need
Every token in the context window is stored in the KV cache. At Q4_K_M, with a 16-bit KV cache, Phi-4 needs:
| Context | KV cache | Total memory | With 8-bit KV |
|---|---|---|---|
| 4K tokens | 0.8 GB | 10.8 GB | 10.3 GB |
| 8K tokens | 1.7 GB | 11.6 GB | 10.8 GB |
| 16K tokens | 3.4 GB | 13.3 GB | 11.6 GB |
2 x 40 layers x 10 KV heads x 128 dims x 2 bytes = 204,800 bytes (200 KiB) per token, about 1.7GB at 8K and 3.4GB at the full 16K context.
Compatibility
Every GPU that runs Phi-4
Each GPU at the highest-quality quantization that fits with 8K context, fastest first. Speeds are single-user estimates.
| GPU | Memory | Quant | Speed | Fit | Price |
|---|---|---|---|---|---|
| RTX PRO 6000 | 96GB | Q8_0 | ~84 tok/s | Runs well | ~$16,000 |
| RTX 5090 | 32GB | Q8_0 | ~84 tok/s | Runs well | ~$4,300 |
| RTX 5070 | 12GB | Q3_K_M | ~69 tok/s | Runs well | ~$800 |
| RTX 5080 | 16GB | Q5_K_M | ~67 tok/s | Runs well | ~$1,600 |
| RTX PRO 5000 72GB | 72GB | Q8_0 | ~65 tok/s | Runs well | ~$9,100 |
| RTX 5070 Ti | 16GB | Q5_K_M | ~63 tok/s | Runs well | ~$1,200 |
| RTX 4070 | 12GB | Q3_K_M | ~52 tok/s | Runs well | ~$560 used |
| RTX 4080 Super | 16GB | Q5_K_M | ~52 tok/s | Runs well | ~$1,150 used |
| RTX 4090 | 24GB | Q8_0 | ~49 tok/s | Runs well | ~$2,500 used |
| RTX 4070 Ti Super | 16GB | Q5_K_M | ~48 tok/s | Runs well | ~$900 used |
| RTX 6000 Ada | 48GB | Q8_0 | ~47 tok/s | Runs well | ~$4,750 used |
| RTX 3090 | 24GB | Q8_0 | ~41 tok/s | Runs well | ~$1,050 used |
| M5 Ultra 512GB | 512GB unified | Q8_0 | ~41 tok/s | Runs well | — |
| M5 Ultra 256GB | 256GB unified | Q8_0 | ~41 tok/s | Runs well | ~$9,499 |
| RX 7900 XTX | 24GB | Q8_0 | ~38 tok/s | Runs well | ~$1,050 used |
| RX 9070 XT | 16GB | Q5_K_M | ~37 tok/s | Runs well | ~$800 |
| V100 32GB | 32GB | Q8_0 | ~36 tok/s | Runs well | ~$700 used |
| RTX 3060 12GB | 12GB | Q3_K_M | ~34 tok/s | Runs well | ~$270 used |
| RTX A6000 | 48GB | Q8_0 | ~34 tok/s | Runs well | ~$4,400 used |
| RTX 5060 Ti 16GB | 16GB | Q5_K_M | ~33 tok/s | Runs well | ~$620 |
| Arc B580 | 12GB | Q3_K_M | ~32 tok/s | Runs well | ~$330 |
| MI50 32GB | 32GB | Q8_0 | ~28 tok/s | Runs well | ~$600 used |
| M3 Ultra 512GB | 512GB unified | Q8_0 | ~28 tok/s | Runs well | $9,499 |
| R9700 | 32GB | Q8_0 | ~26 tok/s | Runs well | ~$1,800 |
| M5 Max 128GB | 128GB unified | Q8_0 | ~25 tok/s | Runs well | ~$5,100 |
| RTX 4000 Ada | 20GB | Q6_K | ~23 tok/s | Runs well | $1,250 |
| M4 Max 128GB | 128GB unified | Q8_0 | ~22 tok/s | Runs well | $3,499 |
| RTX 4060 Ti 16GB | 16GB | Q5_K_M | ~21 tok/s | Runs well | ~$380 used |
| Arc Pro B70 | 32GB | Q8_0 | ~21 tok/s | Runs well | ~$1,300 |
| RX 9060 XT 16GB | 16GB | Q5_K_M | ~19 tok/s | Runs well | ~$470 |
| Arc Pro B60 | 24GB | Q8_0 | ~16 tok/s | Runs well | ~$650 |
| Ryzen AI Max+ PRO 495 | 192GB unified | Q8_0 | ~13 tok/s | Runs well | ~$6,800 |
| DGX Spark | 128GB unified | Q8_0 | ~13 tok/s | Runs well | ~$6,950 |
| Ryzen AI Max+ 395 | 128GB unified | Q8_0 | ~12 tok/s | Runs well | ~$3,300 |
| Tesla P40 | 24GB | Q8_0 | ~12 tok/s | Runs well | ~$230 used |
With CPU offload (partial offload, assuming 64GB of system RAM): RTX 5060 (~9.4 tok/s).
Tips
Getting the most out of Phi-4
- Start at Q4_K_M. It keeps nearly all of the quality at about a third of the 16-bit size. Step up to Q5_K_M or Q6_K only if it still fits with the context you need.
- Budget memory for context. The KV cache grows with every token. Quantizing it to 8-bit roughly halves that cost with little quality loss.
- Avoid heavy CPU offload. For a dense model, every token reads every weight. Once a large share of the model sits in system RAM, speed falls off a cliff. A second GPU usually beats offloading.
- Pick the right runtime. Ollama and LM Studio are easiest. llama.cpp gives the most control. vLLM or SGLang are best when several people or agents share one machine.
Want to try different quantizations, context lengths, or GPU counts? Use our VRAM calculator, or read the Phi-4 overview for setup commands and benchmarks.
Frequently asked questions
How much VRAM do I need to run Phi-4?
About 12GB at Q4_K_M with an 8K context window. At Q8_0 (near-lossless) it needs about 18GB, and at Q3_K_M about 9.8GB with some quality loss. Longer contexts add 0.21GB per 1K tokens with a 16-bit KV cache.
Can I run Phi-4 on a 8GB GPU?
Only with partial CPU offload. On a 8GB card like the RTX 5060, about 45% of the weights would sit in system RAM and generation drops to about 9.4 tokens/sec with CPU offload.
Can I run Phi-4 on a Mac?
Yes. With 128GB of unified memory or more (for example the Ryzen AI Max+ 395), macOS can give the GPU enough memory to run it at Q8_0, at about ~12 tokens/sec. MLX builds are usually fastest on Apple silicon.
Can I run Phi-4 on CPU only, without a GPU?
Yes, if you have at least 32GB of system RAM. On a desktop with dual-channel DDR5, expect roughly 4.9 tokens/sec. A GPU is strongly recommended for interactive use.