Phi · 9B to 32B
Phi 4
14.66 billion parameters across 40 layers. At four-bit precision the weights alone come to 8.24 GB, before any conversation is loaded. This one is cache-hungry: a 32k window adds 6.3 GB on top, so context length matters more here than the model size suggests.
About Phi 4
Microsoft's 14B model, released under the MIT license in December 2024. It was trained heavily on synthetic data and is strong at reasoning and math for its size. Its context window is 16k, which is short next to newer models.
Phi 4 is a dense model with 14.66B parameters, built from 40 layers with a hidden size of 5120. Its repository was published on 2024-12-11. The license listed on Hugging Face is mit. It uses grouped-query attention with 10 KV heads for 40 query heads, which cuts the cache to 25% of what full attention would need.
Memory needed, by quantization
At an 8k context window. Quality is our rough ranking of how much the compression costs you: anything at or above Q5 is hard to tell apart from the original in normal use.
| Quantization | File size | KV cache | Total VRAM | Quality | What it costs you |
|---|---|---|---|---|---|
| Q8_0 | 14.51 GB | 1.56 GB | 16 GB | 99% | Lossless in practice. Use it when the memory is there. |
| Q6_K | 11.2 GB | 1.56 GB | 12.81 GB | 98% | Very close to Q8 for two thirds of the size. |
| Q5_K_M | 9.71 GB | 1.56 GB | 11.37 GB | 96% | The quality-first choice when Q6 will not fit. |
| Q4_K_M | 8.24 GB | 1.56 GB | 9.96 GB | 93% | The default. Best size-to-quality ratio for local use. |
| IQ4_XS | 7.25 GB | 1.56 GB | 9 GB | 91% | Importance-matrix 4-bit. Q4_K_S quality, smaller file. |
| Q3_K_M | 6.67 GB | 1.56 GB | 8.44 GB | 86% | Degradation starts to show. A way to fit one size up. |
| IQ3_XXS | 5.22 GB | 1.56 GB | 7.04 GB | 79% | Aggressive. Only worth it on very large models. |
What a longer conversation costs
Same model at Q4_K_M, only the context window changes. The cache grows in a straight line with every token in the window.
| Context | KV cache | Total VRAM | Fits in 8 GB | Fits in 12 GB | Fits in 24 GB |
|---|---|---|---|---|---|
| 4k | 0.78 GB | 9.18 GB | no | yes | yes |
| 8k | 1.56 GB | 9.96 GB | no | yes | yes |
| 16k | 3.13 GB | 11.52 GB | no | no | yes |
The "fits" columns allow for the roughly 0.8 GB Windows keeps for the desktop.
Which hardware runs Phi 4
93 of 118 consumer devices run it at a quantization worth using, at an 8k context window. Another 11 can load it only by compressing the weights far enough to damage the model, marked with a warning below.
| Device | Memory | Verdict | Quantization | Used | Speed |
|---|---|---|---|---|---|
| GeForce RTX 5090 NVIDIA | 32 GB | Runs great | Q8_0 | 16 GB | 84.7 t/s |
| RTX PRO 6000 Blackwell NVIDIA | 96 GB | Runs great | Q8_0 | 16 GB | 84.7 t/s |
| GeForce RTX 4090 NVIDIA | 24 GB | Runs great | Q8_0 | 16 GB | 47.7 t/s |
| GeForce RTX 3090 Ti NVIDIA | 24 GB | Runs great | Q8_0 | 16 GB | 47.7 t/s |
| RTX 6000 Ada Generation NVIDIA | 48 GB | Runs great | Q8_0 | 16 GB | 45.4 t/s |
| GeForce RTX 3090 NVIDIA | 24 GB | Runs great | Q8_0 | 16 GB | 44.3 t/s |
| GeForce RTX 5090 Laptop NVIDIA | 24 GB | Runs great | Q8_0 | 16 GB | 42.4 t/s |
| Radeon RX 7900 XTX AMD | 24 GB | Runs great | Q8_0 | 16 GB | 38.5 t/s |
| RTX A6000 NVIDIA | 48 GB | Runs great | Q8_0 | 16 GB | 36.3 t/s |
| RTX A5000 NVIDIA | 24 GB | Runs great | Q8_0 | 16 GB | 36.3 t/s |
| Apple M3 Ultra 96GB Apple | 96 GB | Runs great | Q8_0 | 16.48 GB | 35.8 t/s |
| Apple M3 Ultra 256GB Apple | 256 GB | Runs great | Q8_0 | 16.48 GB | 35.8 t/s |
| Apple M3 Ultra 512GB Apple | 512 GB | Runs great | Q8_0 | 16.48 GB | 35.8 t/s |
| Apple M1 Ultra 64GB Apple | 64 GB | Runs great | Q8_0 | 16.48 GB | 35 t/s |
| Apple M1 Ultra 128GB Apple | 128 GB | Runs great | Q8_0 | 16.48 GB | 35 t/s |
| Apple M2 Ultra 64GB Apple | 64 GB | Runs great | Q8_0 | 16.48 GB | 35 t/s |
| Apple M2 Ultra 128GB Apple | 128 GB | Runs great | Q8_0 | 16.48 GB | 35 t/s |
| Apple M2 Ultra 192GB Apple | 192 GB | Runs great | Q8_0 | 16.48 GB | 35 t/s |
| Radeon PRO W7900 AMD | 48 GB | Runs great | Q8_0 | 16 GB | 34.6 t/s |
| Radeon RX 7900 XT AMD | 20 GB | Runs great | Q8_0 | 16 GB | 32.1 t/s |
| Apple M4 Max 36GB Apple | 36 GB | Runs well | Q8_0 | 16.48 GB | 23.9 t/s |
| Apple M4 Max 48GB Apple | 48 GB | Runs well | Q8_0 | 16.48 GB | 23.9 t/s |
| Apple M4 Max 64GB Apple | 64 GB | Runs well | Q8_0 | 16.48 GB | 23.9 t/s |
| Apple M4 Max 128GB Apple | 128 GB | Runs well | Q8_0 | 16.48 GB | 23.9 t/s |
| GeForce RTX 5080 NVIDIA | 16 GB | Runs great | Q6_K | 12.81 GB | 57.1 t/s |
Show 35 more devices
| GeForce RTX 5070 Ti NVIDIA | 16 GB | Runs great | Q6_K | 12.81 GB | 53.3 t/s |
| GeForce RTX 5080 Laptop NVIDIA | 16 GB | Runs great | Q6_K | 12.81 GB | 45.7 t/s |
| GeForce RTX 4080 SUPER NVIDIA | 16 GB | Runs great | Q6_K | 12.81 GB | 43.8 t/s |
| GeForce RTX 4080 NVIDIA | 16 GB | Runs great | Q6_K | 12.81 GB | 42.7 t/s |
| GeForce RTX 4070 Ti SUPER NVIDIA | 16 GB | Runs great | Q6_K | 12.81 GB | 40 t/s |
| GeForce RTX 4090 Laptop NVIDIA | 16 GB | Runs great | Q6_K | 12.81 GB | 34.3 t/s |
| Radeon RX 9070 XT AMD | 16 GB | Runs great | Q6_K | 12.81 GB | 32.5 t/s |
| Radeon RX 9070 AMD | 16 GB | Runs great | Q6_K | 12.81 GB | 32.5 t/s |
| Radeon RX 7800 XT AMD | 16 GB | Runs great | Q6_K | 12.81 GB | 31.5 t/s |
| Radeon RX 7900 GRE AMD | 16 GB | Runs great | Q6_K | 12.81 GB | 29.1 t/s |
| GeForce RTX 5060 Ti 16GB NVIDIA | 16 GB | Runs great | Q6_K | 12.81 GB | 26.7 t/s |
| RTX A4000 NVIDIA | 16 GB | Runs great | Q6_K | 12.81 GB | 26.7 t/s |
| Radeon RX 6900 XT AMD | 16 GB | Runs great | Q6_K | 12.81 GB | 25.8 t/s |
| Radeon RX 6800 AMD | 16 GB | Runs great | Q6_K | 12.81 GB | 25.8 t/s |
| Arc A770 16GB Intel | 16 GB | Runs great | Q6_K | 12.81 GB | 25.7 t/s |
| Apple M1 Max 32GB Apple | 32 GB | Runs well | Q8_0 | 16.48 GB | 17.5 t/s |
| Apple M1 Max 64GB Apple | 64 GB | Runs well | Q8_0 | 16.48 GB | 17.5 t/s |
| Apple M2 Max 32GB Apple | 32 GB | Runs well | Q8_0 | 16.48 GB | 17.5 t/s |
| Apple M2 Max 64GB Apple | 64 GB | Runs well | Q8_0 | 16.48 GB | 17.5 t/s |
| Apple M2 Max 96GB Apple | 96 GB | Runs well | Q8_0 | 16.48 GB | 17.5 t/s |
| Apple M3 Max 36GB Apple | 36 GB | Runs well | Q8_0 | 16.48 GB | 17.5 t/s |
| Apple M3 Max 48GB Apple | 48 GB | Runs well | Q8_0 | 16.48 GB | 17.5 t/s |
| Apple M3 Max 64GB Apple | 64 GB | Runs well | Q8_0 | 16.48 GB | 17.5 t/s |
| Apple M3 Max 96GB Apple | 96 GB | Runs well | Q8_0 | 16.48 GB | 17.5 t/s |
| Apple M3 Max 128GB Apple | 128 GB | Runs well | Q8_0 | 16.48 GB | 17.5 t/s |
| GeForce RTX 4060 Ti 16GB NVIDIA | 16 GB | Runs well | Q6_K | 12.81 GB | 17.1 t/s |
| GeForce RTX 3080 Ti NVIDIA | 12 GB | Runs great | Q4_K_M | 9.96 GB | 70.5 t/s |
| GeForce RTX 3080 12GB NVIDIA | 12 GB | Runs great | Q4_K_M | 9.96 GB | 70.5 t/s |
| GeForce RTX 5070 NVIDIA | 12 GB | Runs great | Q4_K_M | 9.96 GB | 52 t/s |
| GeForce RTX 2080 Ti NVIDIA | 11 GB | Just fits | Q4_K_M | 9.96 GB | 47.6 t/s |
| GeForce RTX 4070 Ti NVIDIA | 12 GB | Runs great | Q4_K_M | 9.96 GB | 39 t/s |
| GeForce RTX 4070 SUPER NVIDIA | 12 GB | Runs great | Q4_K_M | 9.96 GB | 39 t/s |
| GeForce RTX 4070 NVIDIA | 12 GB | Runs great | Q4_K_M | 9.96 GB | 39 t/s |
| GeForce GTX 1080 Ti NVIDIA | 11 GB | Just fits | Q4_K_M | 9.96 GB | 37.4 t/s |
| GeForce RTX 4080 Laptop NVIDIA | 12 GB | Runs great | Q4_K_M | 9.96 GB | 33.4 t/s |
Can my GPU run Phi 4?
A direct answer for popular devices, each with the settings we would use.
How it compares
Other models in the same family or of a similar size, on a 12 GB card and a 24 GB card, each at the best quantization that fits.
| Model | Weights at 4 bits | GeForce RTX 3060 12GB | GeForce RTX 4090 |
|---|---|---|---|
| Phi 4 | 8.2 GB | 27.8 t/s at Q4_K_M | 47.7 t/s at Q8_0 |
| Phi 4 Mini Instruct | 2.2 GB | 55.2 t/s at Q8_0 | 154.5 t/s at Q8_0 |
| Phi 3 Mini 4k Instruct | 2.1 GB | 59.8 t/s at Q8_0 | 167.3 t/s at Q8_0 |
| Phi 3.5 MoE Instruct | 23.5 GB | 13.1 t/s at Q4_K_M | 100.6 t/s at IQ4_XS |
| Qwen3.5 9B | 5.4 GB | 30 t/s at Q8_0 | 83.9 t/s at Q8_0 |
| GPT OSS 20B | 11.3 GB | 65.7 t/s at MXFP4 | 180.1 t/s at MXFP4 |
What to buy to run Phi 4
The cheapest hardware that runs it at a quantization worth using and a speed you would not resent, at an 8k context window.
Cheapest that works
GeForce RTX 2060 12GB
$160 used
Q4_K_M · uses 9.96 GB of its 12 GB · about 26 tokens/s
Best value
GeForce RTX 3060 12GB
$200 used
Q4_K_M · about 27.8 tokens/s · 360 GB/s
| Device | Price | Memory | Runs it at | Speed | |
|---|---|---|---|---|---|
| GeForce GTX 1080 Ti NVIDIA | $150 used | 11 GB | Q4_K_M | 37.4 t/s | Check price |
| GeForce RTX 2060 12GB NVIDIA | $160 used | 12 GB | Q4_K_M | 26 t/s | Check price |
| GeForce RTX 2080 Ti NVIDIA | $250 used | 11 GB | Q4_K_M | 47.6 t/s | Check price |
| Arc A770 16GB Intel | $280 used | 16 GB | Q6_K | 25.7 t/s | Check price |
| GeForce RTX 3080 10GB NVIDIA | $350 used | 10 GB | IQ4_XS | 65.3 t/s | Check price |
| GeForce RTX 3080 12GB NVIDIA | $400 used | 12 GB | Q4_K_M | 70.5 t/s | Check price |
| Radeon RX 7900 XT AMD | $650 | 20 GB | Q8_0 | 32.1 t/s | Check price |
| GeForce RTX 3090 NVIDIA | $700 used | 24 GB | Q8_0 | 44.3 t/s | Check price |
Indicative prices reviewed 2026-09-01; used prices are marketplace typical. Options that cost more than a cheaper one with no more memory and no more speed are hidden. Price links are Amazon affiliate links. Change the standard, the context window or the budget in the buying tool.
Where it sits in the catalog
Source: microsoft/Phi-4 on Hugging Face. Downloaded 632k times in the last month. Published 2024-12-11. Architecture figures are read from the repository's own configuration file, so they move when the model does.
Questions people ask about Phi 4
How much VRAM does Phi 4 need?
At Q4_K_M with an 8k context window, Phi 4 needs about 9.96 GB: 7.95 GB of weights, 1.56 GB of KV cache and 0.44 GB of overhead. At the highest-quality file it needs about 16 GB.
What is the cheapest GPU that runs Phi 4?
The smallest consumer device in our catalog that fits it is the GeForce RTX 3080 10GB with 10 GB, running it at IQ4_XS at roughly 65.3 tokens per second.
How much does a longer context window add for Phi 4?
On the memory side, a 32k window adds 6.3 GB of KV cache on top of the weights, and the cache grows in a straight line with every token. The longest window the model supports natively is 16k tokens.