Phi · MIT · HF: microsoft/phi-4
Phi-4 14B — VRAM, KV Cache & Hardware Matrix
Serverless API Rate
$0.07 in / $0.14 out per 1M
Parameters
14.7B Dense
Attention
GQA (10 KV / 40 Q)
Layers × Hidden
40 × 5120
Max Native Context
16k tokens
MoE Experts
N/A (Dense)
Quantization Memory Footprint
| Quantization | BPW | Weights Only | Total @ 8k Ctx | Total @ 32k Ctx | Quality |
|---|---|---|---|---|---|
| FP16 | 16.00 | 27.93 GiB | 30.94 GiB | 32.51 GiB | Lossless Reference |
| Q8_0 | 8.50 | 14.84 GiB | 17.85 GiB | 19.42 GiB | Near-Lossless |
| Q6_K | 6.57 | 11.47 GiB | 14.48 GiB | 16.05 GiB | High Precision |
| Q5_K_M | 5.69 | 9.93 GiB | 12.95 GiB | 14.51 GiB | Balanced High |
| Q4_K_M | 4.85 | 8.47 GiB | 11.48 GiB | 13.04 GiB | Recommended 4-bit |
| MLX_4BIT | 4.50 | 7.85 GiB | 10.87 GiB | 12.43 GiB | Apple Silicon Optimized |
| IQ4_XS | 4.25 | 7.42 GiB | 10.43 GiB | 12.00 GiB | Compact 4-bit |
| Q3_K_M | 3.91 | 6.83 GiB | 9.84 GiB | 11.40 GiB | Aggressive Compression |
Hardware Compatibility & Speed Leaderboard (Q4_K_M @ 16k Context)
| Hardware | Status | Total / Limit | Decode | TTFT (4k) | Max Safe Ctx | Inspect |
|---|---|---|---|---|---|---|
| Cloud NVIDIA H200 SXM (141 GB) | COMFORTABLE | 13.0 / 140.4 GiB | 324.0 tok/s | 0.21s | 16k | Simulate → |
| Cloud NVIDIA H100 SXM (80 GB) | COMFORTABLE | 13.0 / 79.4 GiB | 226.1 tok/s | 0.21s | 16k | Simulate → |
| Cloud NVIDIA A100 SXM (80 GB) | COMFORTABLE | 13.0 / 79.4 GiB | 134.3 tok/s | 0.64s | 16k | Simulate → |
| NVIDIA RTX PRO 6000 Blackwell (96 GB) | COMFORTABLE | 13.0 / 95.0 GiB | 121.0 tok/s | 0.78s | 16k | Simulate → |
| NVIDIA RTX 5090 (32 GB) | COMFORTABLE | 13.0 / 31.0 GiB | 118.1 tok/s | 0.97s | 16k | Simulate → |
| NVIDIA RTX 4090 (24 GB) | COMFORTABLE | 13.0 / 23.0 GiB | 66.4 tok/s | 1.22s | 16k | Simulate → |
| NVIDIA RTX 5080 (16 GB) | COMFORTABLE | 13.0 / 15.0 GiB | 63.3 tok/s | 1.80s | 16k | Simulate → |
| NVIDIA RTX 3090 (24 GB Used) | COMFORTABLE | 13.0 / 23.0 GiB | 60.2 tok/s | 2.93s | 16k | Simulate → |
| NVIDIA RTX 5070 Ti (16 GB) | COMFORTABLE | 13.0 / 15.0 GiB | 59.0 tok/s | 2.30s | 16k | Simulate → |
| AMD Radeon RX 7900 XTX (24 GB) | COMFORTABLE | 13.0 / 23.0 GiB | 58.6 tok/s | 2.03s | 16k | Simulate → |
| Cloud NVIDIA L40S (48 GB) | COMFORTABLE | 13.0 / 47.4 GiB | 56.9 tok/s | 0.57s | 16k | Simulate → |
| Apple Mac Studio M3 Ultra (256 GB) | COMFORTABLE | 12.8 / 199.7 GiB | 51.3 tok/s | 3.82s | 16k | Simulate → |
| Apple Mac Studio M3 Ultra (512 GB) | COMFORTABLE | 12.8 / 399.4 GiB | 51.3 tok/s | 3.82s | 16k | Simulate → |
| 2x NVIDIA RTX 3090 Rig (48 GB) | COMFORTABLE | 13.0 / 46.6 GiB | 50.5 tok/s | 3.59s | 16k | Simulate → |
| Apple Mac Studio M2 Ultra (192 GB) | COMFORTABLE | 12.8 / 149.8 GiB | 50.1 tok/s | 4.21s | 16k | Simulate → |
| Apple Mac Studio M4 Max (64 GB) | COMFORTABLE | 12.8 / 49.9 GiB | 33.8 tok/s | 5.80s | 16k | Simulate → |
| Apple Mac Studio M4 Max (128 GB) | COMFORTABLE | 12.8 / 99.8 GiB | 33.8 tok/s | 5.80s | 16k | Simulate → |
| NVIDIA RTX 5060 Ti (16 GB) | COMFORTABLE | 13.0 / 15.0 GiB | 28.8 tok/s | 4.37s | 16k | Simulate → |
| Apple MacBook Pro M3 Max (64 GB) | COMFORTABLE | 12.8 / 49.9 GiB | 24.4 tok/s | 7.76s | 16k | Simulate → |
| Apple MacBook Pro M3 Max (128 GB) | COMFORTABLE | 12.8 / 99.8 GiB | 24.4 tok/s | 7.76s | 16k | Simulate → |
| Apple Mac Mini M4 Pro (48 GB) | COMFORTABLE | 12.8 / 37.4 GiB | 16.2 tok/s | 11.96s | 16k | Simulate → |
| Apple Mac Mini M4 Pro (64 GB) | COMFORTABLE | 12.8 / 49.9 GiB | 16.2 tok/s | 11.96s | 16k | Simulate → |
| Ryzen AI Max+ 395 Strix Halo (128 GB) | COMFORTABLE | 12.8 / 96.0 GiB | 15.2 tok/s | 4.32s | 16k | Simulate → |
| Ryzen AI Max+ 395 Strix Halo (64 GB) | COMFORTABLE | 12.8 / 48.0 GiB | 15.2 tok/s | 4.32s | 16k | Simulate → |
| Apple Mac Mini M4 (24 GB) | COMFORTABLE | 12.8 / 18.0 GiB | 6.9 tok/s | 26.47s | 16k | Simulate → |
| Ryzen AI 9 HX 370 Mini-PC (64 GB) | COMFORTABLE | 12.8 / 48.0 GiB | 6.8 tok/s | 12.33s | 16k | Simulate → |
| NVIDIA RTX 3060 (12 GB) | OOM | 13.0 / 11.0 GiB | OOM | — | 6k | Simulate → |
| Apple Mac Mini M4 (16 GB) | OOM | 12.8 / 10.0 GiB | OOM | — | 9k | Simulate → |
Interactive Phi-4 14B Simulator
14.7B Dense · GQA
1792 GB/s · 250 TFLOPS
3. Weight Quantization
4.85 bpw
4. KV Cache
32,768 tok
Compute Roofline
COMFORTABLE FIT Memory-Bandwidth Bound
Max Safe Ctx: 16k
Memory Allocation Footprint 16.17 / 95.00 GiB usable
Weights: 8.47 GiB
KV Cache: 6.25 GiB
Overhead: 1.45 GiB
Headroom: 78.83 GiB
Decode Speed
95.3
tok/s (batch=1)
Prefill Speed
5470
prompt tok/s
TTFT Latency
0.78s
for 4,096 tok
Active Params
14.7B
of 14.7B total
llama.cpp Server Command
llama-server -hf microsoft/phi-4:Q4_K_M -ngl 99 -fa on -c 32768 --port 8080| Context | KV Cache | Total VRAM | Decode | Status |
|---|---|---|---|---|
| 4k | 0.78 GiB | 10.35 GiB | 151.6 tok/s | COMFORTABLE |
| 8k | 1.56 GiB | 11.48 GiB | 139.8 tok/s | COMFORTABLE |
| 16k | 3.13 GiB | 13.04 GiB | 121.0 tok/s | COMFORTABLE |
Buy-vs-Cloud-vs-API Monthly TCO Break-Even
USE SERVERLESS APIOwn Hardware/mo
$163.99
46.3h active + power
Cloud GPU/mo
$98.46
@$1.85/hr spot/pod
Serverless API/mo
$5.60
API cheaper
Recommended Hardware & Cloud Deployment
Contextual to fit statusEngineering FAQ — Phi-4 14B
How much VRAM does Phi-4 14B require in Q4_K_M?
Phi-4 14B weights occupy 8.47 GiB in Q4_K_M. Including 16k FP16 KV cache and runtime buffers, total memory is 13.04 GiB.
What is the cheapest local hardware that runs Phi-4 14B?
NVIDIA RTX PRO 6000 Blackwell (96 GB) (96 GB) runs Phi-4 14B Q4_K_M at ~121.0 tok/s decode and 0.78s TTFT.
How does Phi-4 14B scale KV cache at long context?
Phi-4 14B uses GQA with 10 KV heads across 40 layers (128 head_dim).