Phi-4 14B VRAM Requirements
Phi-4 14B needs about 10.6 GB of VRAM at Q4_K_M with an 8K context: 8.3 GB for weights, 1.6 GB for the KV cache and 0.8 GB of runtime overhead. At Q8_0 it needs 16.9 GB and at full FP16 precision 29.7 GB. The smallest GPU that runs it fully is the Apple M4 16GB.
10.6 GB VRAM
Phi-4 14B at Q4_K_M with 8K context and the KV cache in FP16.
- Weights 8.3 GB
- KV cache 1.6 GB
- Runtime overhead 0.8 GB
Which GPUs can run it
- RTX 3060 12GBFits~24 tok/s
- RTX 4060 Ti 16GBFits~19 tok/s
- RTX 3090 24GBFits~61 tok/s
- Radeon RX 7900 XTX 24GBFits~53 tok/s
- RTX 4090 24GBFits~65 tok/s
- RTX 5090 32GBFits~114 tok/s
- RTX 6000 Ada 48GBFits~62 tok/s
- A100 80GBFits~129 tok/s
- H100 80GB SXMFits~205 tok/s
- Ryzen AI Max+ 395 128GBFits~16 tok/s
- Apple M4 Max 128GBFits~36 tok/s
- Apple M5 Max 128GBFits~40 tok/s
- RTX PRO 6000 Blackwell 96GBFits~114 tok/s
- NVIDIA DGX Spark 128GBFits~17 tok/s
- Apple M3 Ultra 512GBFits~53 tok/s
Phi-4 14B VRAM by quantization
Total memory with an FP16 KV cache. "Smallest tier" is the smallest standard VRAM size with a 6% safety margin at 8K.
| Quantization | Bits/weight | Weights | Total at 8K | Total at 16K | Smallest tier |
|---|---|---|---|---|---|
| Q4_K_M | 4.85 | 8.3 GB | 10.6 GB | 12.2 GB | 12 GB |
| Q5_K_M | 5.69 | 9.7 GB | 12.1 GB | 13.6 GB | 16 GB |
| Q6_K | 6.59 | 11.3 GB | 13.6 GB | 15.2 GB | 16 GB |
| Q8_0 | 8.50 | 14.5 GB | 16.9 GB | 18.4 GB | 20 GB |
| FP16 / BF16 | 16.00 | 27.4 GB | 29.7 GB | 31.3 GB | 32 GB |
| Q3_K_M | 3.91 | 6.7 GB | 9.0 GB | 10.6 GB | 10 GB |
| Q2_K | 3.00 | 5.1 GB | 7.5 GB | 9.0 GB | 8 GB |
VRAM by context length
At Q4_K_M. The KV cache grows linearly with context: 0.20 GB per 1,000 tokens in FP16.
| Context | KV cache (FP16) | KV cache (Q8_0) | Total, FP16 cache | Total, Q8_0 cache |
|---|---|---|---|---|
| 4K tokens | 0.8 GB | 0.4 GB | 9.9 GB | 9.5 GB |
| 8K tokens | 1.6 GB | 0.8 GB | 10.6 GB | 9.9 GB |
| 16K tokens | 3.1 GB | 1.7 GB | 12.2 GB | 10.7 GB |
GPUs that can run Phi-4 14B
At Q4_K_M with 8K context. "Best quant" is the highest-quality format that fits fully; "max context" is at Q4_K_M.
| GPU | Usable memory | Status | Best quant | Max context | Speed |
|---|---|---|---|---|---|
| Apple M4 16GB | 10.7 GB | Tight fit | IQ4_XS | 5K | ~8.0 tok/s |
| RTX 2080 Ti 11GB | 11.0 GB | Tight fit | IQ4_XS | 6K | ~40 tok/s |
| RTX 2060 12GB | 12.0 GB | Fits | Q4_K_M | 11K | ~22 tok/s |
| RTX 3060 12GB | 12.0 GB | Fits | Q4_K_M | 11K | ~24 tok/s |
| Radeon RX 6700 XT 12GB | 12.0 GB | Fits | Q4_K_M | 11K | ~21 tok/s |
| Radeon RX 7700 XT 12GB | 12.0 GB | Fits | Q4_K_M | 11K | ~24 tok/s |
| Intel Arc B580 12GB | 12.0 GB | Fits | Q4_K_M | 11K | ~23 tok/s |
| RTX 4070 12GB | 12.0 GB | Fits | Q4_K_M | 11K | ~33 tok/s |
| RTX 4070 Super 12GB | 12.0 GB | Fits | Q4_K_M | 11K | ~33 tok/s |
| RTX 4070 Ti 12GB | 12.0 GB | Fits | Q4_K_M | 11K | ~33 tok/s |
| RTX 5070 12GB | 12.0 GB | Fits | Q4_K_M | 11K | ~44 tok/s |
| RTX 3080 12GB | 12.0 GB | Fits | Q4_K_M | 11K | ~59 tok/s |
| Apple M4 Pro 24GB | 16.0 GB | Fits | Q6_K | 16K | ~18 tok/s |
| RTX 4060 Ti 16GB | 16.0 GB | Fits | Q6_K | 16K | ~19 tok/s |
| RTX 3080 10GB | 10.0 GB | 88% on GPU | Q3_K_M | 1K | ~22 tok/s |
| Intel Arc B570 10GB | 10.0 GB | 88% on GPU | Q3_K_M | 1K | ~14 tok/s |
| RTX 3050 8GB | 8.0 GB | 71% on GPU | Q2_K | – | ~8.6 tok/s |
| RTX 3060 Ti 8GB | 8.0 GB | 71% on GPU | Q2_K | – | ~11 tok/s |
| RTX 3070 8GB | 8.0 GB | 71% on GPU | Q2_K | – | ~11 tok/s |
| RTX 3070 Ti 8GB | 8.0 GB | 71% on GPU | Q2_K | – | ~12 tok/s |
About Phi-4 14B
A 14B dense model focused on reasoning and maths. Its 16K context limit keeps total memory predictable: Q4_K_M at full context needs about 12.2 GB, or 10.7 GB with a Q8_0 KV cache.
| Spec | Value |
|---|---|
| Parameters | 14.7B |
| Layers | 40 |
| Hidden size | 5120 |
| Attention heads / KV heads | 40 / 10 |
| Head dimension | 128 |
| Vocabulary | 100,352 |
| Max context (used here) | 16K tokens |
| KV cache per token (FP16) | 200 KB |
| Released | 2024-12 |
| License | MIT |
Phi-4 14B VRAM: common questions
Does Phi-4 fit on a 12GB GPU?
Yes with a little tuning. At Q4_K_M it needs about 12.2 GB at its full 16K context, so on 12 GB use a Q8_0 KV cache (about 10.7 GB) or keep the context near 8K.
How much VRAM does Phi-4 14B need?
About 10.6 GB at Q4_K_M with an 8K context, 16.9 GB at Q8_0 and 29.7 GB at FP16. Longer contexts add 0.20 GB per 1,000 tokens with an FP16 cache.
What is the cheapest GPU that can run Phi-4 14B?
The smallest listed device that fits it fully at Q4_K_M with 8K context is the Apple M4 16GB (10.7 GB usable).