DeepSeek · MIT · HF: deepseek-ai/DeepSeek-R1-Distill-Qwen-14B
DeepSeek-R1 Distill Qwen 14B — VRAM, KV Cache & Hardware Matrix
Serverless API Rate
$0.06 in / $0.20 out per 1M
Parameters
14.8B Dense
Attention
GQA (8 KV / 40 Q)
Layers × Hidden
48 × 5120
Max Native Context
128k tokens
MoE Experts
N/A (Dense)
Quantization Memory Footprint
| Quantization | BPW | Weights Only | Total @ 8k Ctx | Total @ 32k Ctx | Quality |
|---|---|---|---|---|---|
| FP16 | 16.00 | 28.12 GiB | 31.07 GiB | 35.57 GiB | Lossless Reference |
| Q8_0 | 8.50 | 14.94 GiB | 17.89 GiB | 22.39 GiB | Near-Lossless |
| Q6_K | 6.57 | 11.55 GiB | 14.50 GiB | 19.00 GiB | High Precision |
| Q5_K_M | 5.69 | 10.00 GiB | 12.95 GiB | 17.45 GiB | Balanced High |
| Q4_K_M | 4.85 | 8.52 GiB | 11.48 GiB | 15.98 GiB | Recommended 4-bit |
| MLX_4BIT | 4.50 | 7.91 GiB | 10.86 GiB | 15.36 GiB | Apple Silicon Optimized |
| IQ4_XS | 4.25 | 7.47 GiB | 10.42 GiB | 14.92 GiB | Compact 4-bit |
| Q3_K_M | 3.91 | 6.87 GiB | 9.82 GiB | 14.32 GiB | Aggressive Compression |
Hardware Compatibility & Speed Leaderboard (Q4_K_M @ 32k Context)
| Hardware | Status | Total / Limit | Decode | TTFT (4k) | Max Safe Ctx | Inspect |
|---|---|---|---|---|---|---|
| Cloud NVIDIA H200 SXM (141 GB) | COMFORTABLE | 16.0 / 140.4 GiB | 258.6 tok/s | 0.22s | 128k | Simulate → |
| Cloud NVIDIA H100 SXM (80 GB) | COMFORTABLE | 16.0 / 79.4 GiB | 180.4 tok/s | 0.22s | 128k | Simulate → |
| Cloud NVIDIA A100 SXM (80 GB) | COMFORTABLE | 16.0 / 79.4 GiB | 107.2 tok/s | 0.65s | 128k | Simulate → |
| NVIDIA RTX PRO 6000 Blackwell (96 GB) | COMFORTABLE | 16.0 / 95.0 GiB | 96.5 tok/s | 0.80s | 128k | Simulate → |
| NVIDIA RTX 5090 (32 GB) | COMFORTABLE | 16.0 / 31.0 GiB | 94.2 tok/s | 0.99s | 112k | Simulate → |
| NVIDIA RTX 4090 (24 GB) | COMFORTABLE | 16.0 / 23.0 GiB | 53.0 tok/s | 1.24s | 69k | Simulate → |
| NVIDIA RTX 3090 (24 GB Used) | COMFORTABLE | 16.0 / 23.0 GiB | 48.0 tok/s | 2.98s | 69k | Simulate → |
| AMD Radeon RX 7900 XTX (24 GB) | COMFORTABLE | 16.0 / 23.0 GiB | 46.8 tok/s | 2.06s | 69k | Simulate → |
| Cloud NVIDIA L40S (48 GB) | COMFORTABLE | 16.0 / 47.4 GiB | 45.4 tok/s | 0.58s | 128k | Simulate → |
| Apple Mac Studio M3 Ultra (256 GB) | COMFORTABLE | 15.8 / 199.7 GiB | 41.0 tok/s | 3.88s | 128k | Simulate → |
| Apple Mac Studio M3 Ultra (512 GB) | COMFORTABLE | 15.8 / 399.4 GiB | 41.0 tok/s | 3.88s | 128k | Simulate → |
| 2x NVIDIA RTX 3090 Rig (48 GB) | COMFORTABLE | 16.0 / 46.6 GiB | 40.3 tok/s | 3.65s | 128k | Simulate → |
| Apple Mac Studio M2 Ultra (192 GB) | COMFORTABLE | 15.8 / 149.8 GiB | 40.0 tok/s | 4.29s | 128k | Simulate → |
| Apple Mac Studio M4 Max (64 GB) | COMFORTABLE | 15.8 / 49.9 GiB | 27.0 tok/s | 5.90s | 128k | Simulate → |
| Apple Mac Studio M4 Max (128 GB) | COMFORTABLE | 15.8 / 99.8 GiB | 27.0 tok/s | 5.90s | 128k | Simulate → |
| Apple MacBook Pro M3 Max (64 GB) | COMFORTABLE | 15.8 / 49.9 GiB | 19.5 tok/s | 7.90s | 128k | Simulate → |
| Apple MacBook Pro M3 Max (128 GB) | COMFORTABLE | 15.8 / 99.8 GiB | 19.5 tok/s | 7.90s | 128k | Simulate → |
| Apple Mac Mini M4 Pro (48 GB) | COMFORTABLE | 15.8 / 37.4 GiB | 12.9 tok/s | 12.17s | 128k | Simulate → |
| Apple Mac Mini M4 Pro (64 GB) | COMFORTABLE | 15.8 / 49.9 GiB | 12.9 tok/s | 12.17s | 128k | Simulate → |
| Ryzen AI Max+ 395 Strix Halo (128 GB) | COMFORTABLE | 15.8 / 96.0 GiB | 12.2 tok/s | 4.39s | 128k | Simulate → |
| Ryzen AI Max+ 395 Strix Halo (64 GB) | COMFORTABLE | 15.8 / 48.0 GiB | 12.2 tok/s | 4.39s | 128k | Simulate → |
| Apple Mac Mini M4 (24 GB) | COMFORTABLE | 15.8 / 18.0 GiB | 5.5 tok/s | 26.92s | 52k | Simulate → |
| Ryzen AI 9 HX 370 Mini-PC (64 GB) | COMFORTABLE | 15.8 / 48.0 GiB | 5.4 tok/s | 12.55s | 128k | Simulate → |
| NVIDIA RTX 5080 (16 GB) | OOM | 16.0 / 15.0 GiB | OOM | — | 27k | Simulate → |
| NVIDIA RTX 5070 Ti (16 GB) | OOM | 16.0 / 15.0 GiB | OOM | — | 27k | Simulate → |
| NVIDIA RTX 5060 Ti (16 GB) | OOM | 16.0 / 15.0 GiB | OOM | — | 27k | Simulate → |
| NVIDIA RTX 3060 (12 GB) | OOM | 16.0 / 11.0 GiB | OOM | — | 6k | Simulate → |
| Apple Mac Mini M4 (16 GB) | OOM | 15.8 / 10.0 GiB | OOM | — | 9k | Simulate → |
Interactive DeepSeek-R1 Distill Qwen 14B Simulator
14.8B 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: 128k
Memory Allocation Footprint 15.98 / 95.00 GiB usable
Weights: 8.52 GiB
KV Cache: 6.00 GiB
Overhead: 1.45 GiB
Headroom: 79.02 GiB
Decode Speed
96.5
tok/s (batch=1)
Prefill Speed
5377
prompt tok/s
TTFT Latency
0.80s
for 4,096 tok
Active Params
14.8B
of 14.8B total
llama.cpp Server Command
llama-server -hf deepseek-ai/DeepSeek-R1-Distill-Qwen-14B:Q4_K_M -ngl 99 -fa on -c 32768 --port 8080| Context | KV Cache | Total VRAM | Decode | Status |
|---|---|---|---|---|
| 4k | 0.75 GiB | 10.37 GiB | 151.2 tok/s | COMFORTABLE |
| 8k | 1.50 GiB | 11.48 GiB | 139.9 tok/s | COMFORTABLE |
| 16k | 3.00 GiB | 12.98 GiB | 121.7 tok/s | COMFORTABLE |
| 32k | 6.00 GiB | 15.98 GiB | 96.5 tok/s | COMFORTABLE |
| 64k | 12.00 GiB | 21.98 GiB | 68.3 tok/s | COMFORTABLE |
| 128k | 24.00 GiB | 33.98 GiB | 43.1 tok/s | COMFORTABLE |
Buy-vs-Cloud-vs-API Monthly TCO Break-Even
USE SERVERLESS APIOwn Hardware/mo
$163.94
45.7h active + power
Cloud GPU/mo
$97.33
@$1.85/hr spot/pod
Serverless API/mo
$6.00
API cheaper
Recommended Hardware & Cloud Deployment
Contextual to fit statusEngineering FAQ — DeepSeek-R1 Distill Qwen 14B
How much VRAM does DeepSeek-R1 Distill Qwen 14B require in Q4_K_M?
DeepSeek-R1 Distill Qwen 14B weights occupy 8.52 GiB in Q4_K_M. Including 32k FP16 KV cache and runtime buffers, total memory is 15.98 GiB.
What is the cheapest local hardware that runs DeepSeek-R1 Distill Qwen 14B?
NVIDIA RTX PRO 6000 Blackwell (96 GB) (96 GB) runs DeepSeek-R1 Distill Qwen 14B Q4_K_M at ~96.5 tok/s decode and 0.80s TTFT.
How does DeepSeek-R1 Distill Qwen 14B scale KV cache at long context?
DeepSeek-R1 Distill Qwen 14B uses GQA with 8 KV heads across 48 layers (128 head_dim).