Qwen · Qwen2.5 · 14.8B parameters
Qwen2.5 14B VRAM requirements
Qwen2.5 14B has 48 layers and uses grouped-query attention (8 KV heads). At Q4_K_M the weights come to 8.5 GB, and the best quantisation that fits a 24 GB card is INT8 / W8A8.
Runs comfortably
10.8 GB of 21.8 GB · 50%Qwen2.5 14B at Q4_K_M leaves 11.0 GB spare on a RTX 4090. There is room to raise the context length or move up a quantisation level.
Every quantisation of Qwen2.5 14B on a RTX 4090
Highlighted row is the highest quality that still fits at 8K context.
| Quantisation | bpw | Weights | Total | Verdict | Max ctx | tok/s |
|---|---|---|---|---|---|---|
| FP16 / BF16 | 16.00 | 27.5 GB | 29.8 GB | Won't fit | — | 3.80 |
| INT8 / W8A8 | 8.50 | 14.3 GB | 16.6 GB | Runs comfortably | 36K | 39.8 |
| Q8_0 (GGUF) | 8.50 | 14.3 GB | 16.6 GB | Runs comfortably | 36K | 39.8 |
| FP8 (E4M3) | 8.00 | 13.8 GB | 16.1 GB | Runs comfortably | 38K | 41.2 |
| Q6_K | 6.56 | 11.3 GB | 13.6 GB | Runs comfortably | 51K | 49.5 |
| Q5_K_M | 5.67 | 9.7 GB | 12.1 GB | Runs comfortably | 60K | 56.6 |
| Q5_K_S | 5.52 | 9.5 GB | 11.8 GB | Runs comfortably | 61K | 58.0 |
| AWQ 4-bit | 4.25 | 9.4 GB | 11.8 GB | Runs comfortably | 61K | 58.3 |
| GPTQ 4-bit | 4.25 | 9.4 GB | 11.8 GB | Runs comfortably | 61K | 58.3 |
| MXFP4 | 4.25 | 9.4 GB | 11.8 GB | Runs comfortably | 61K | 58.3 |
| Q4_K_M | 4.85 | 8.5 GB | 10.8 GB | Runs comfortably | 66K | 64.4 |
| Q4_K_S | 4.58 | 8.0 GB | 10.4 GB | Runs comfortably | 69K | 67.4 |
| Q4_0 | 4.55 | 8.0 GB | 10.3 GB | Runs comfortably | 69K | 67.8 |
| IQ4_XS | 4.25 | 7.5 GB | 9.9 GB | Runs comfortably | 71K | 71.5 |
| Q3_K_M | 3.91 | 7.0 GB | 9.3 GB | Runs comfortably | 74K | 76.2 |
| IQ3_M | 3.70 | 6.7 GB | 9.0 GB | Runs comfortably | 76K | 79.4 |
| IQ3_XXS | 3.06 | 5.7 GB | 8.0 GB | Runs comfortably | 81K | 91.2 |
| Q2_K | 2.63 | 5.0 GB | 7.4 GB | Runs comfortably | 85K | 101 |
| IQ2_XXS | 2.06 | 4.2 GB | 6.5 GB | Runs comfortably | 89K | 119 |
| IQ1_M | 1.75 | 3.7 GB | 6.0 GB | Runs comfortably | 92K | 131 |
Qwen2.5 14B on each GPU
Q4_K_M weights at 8K context, single card, monitor attached.
| GPU | VRAM | GB/s | Verdict | Max ctx | tok/s |
|---|---|---|---|---|---|
| H100 SXM 80GB | 80 | 3350 | Runs comfortably | 128K | 237 |
| A100 80GB | 80 | 2039 | Runs comfortably | 128K | 133 |
| RTX 5090 | 32 | 1792 | Runs comfortably | 107K | 124 |
| RTX 5080 | 16 | 960 | Runs comfortably | 26K | 67.6 |
| RTX 4090 | 24 | 1008 | Runs comfortably | 66K | 64.4 |
| RTX 5070 Ti | 16 | 896 | Runs comfortably | 26K | 63.2 |
| RTX 3090 | 24 | 936 | Runs comfortably | 66K | 62.5 |
| Radeon RX 7900 XTX | 24 | 960 | Runs comfortably | 66K | 58.3 |
| L40S | 48 | 864 | Runs comfortably | 128K | 55.4 |
| RTX A6000 | 48 | 768 | Runs comfortably | 128K | 51.4 |
| Mac Studio M3 Ultra 256GB | 256 | 819 | Runs comfortably | 128K | 50.6 |
| RTX 4080 Super | 16 | 736 | Runs comfortably | 26K | 47.3 |
| RTX 4070 Ti Super | 16 | 672 | Runs comfortably | 26K | 43.2 |
| Mac Studio M4 Max 128GB | 128 | 546 | Runs comfortably | 128K | 37.7 |
| RTX 5070 | 12 | 672 | Won't fit | 6K | 34.2 |
| RTX 5060 Ti 16GB | 16 | 448 | Runs comfortably | 26K | 31.9 |
| RTX 4070 Super | 12 | 504 | Won't fit | 6K | 25.3 |
| RTX 4070 | 12 | 504 | Won't fit | 6K | 25.3 |
| RTX 3060 12GB | 12 | 360 | Won't fit | 6K | 20.4 |
| Arc B580 | 12 | 456 | Won't fit | 6K | 19.9 |
| NVIDIA DGX Spark (GB10) | 128 | 273 | Runs comfortably | 128K | 19.5 |
| Mac Mini M4 Pro 48GB | 48 | 273 | Runs comfortably | 128K | 19.0 |
| RTX 4060 Ti 16GB | 16 | 288 | Runs comfortably | 26K | 18.7 |
| Ryzen AI Max+ 395 128GB | 128 | 256 | Runs comfortably | 128K | 14.5 |
| RTX 3080 10GB | 10 | 760 | Won't fit | — | 13.0 |
Architecture
| Parameters | 14.8B |
| Layers | 48 |
| Hidden size | 5120 |
| Attention heads / KV heads | 40 / 8 |
| Head dimension | 128 |
| Vocabulary | 152,064 |
| Trained context | 128K |
| KV cache per 1K tokens | 0 GB |
| Hugging Face | Qwen/Qwen2.5-14B-Instruct |
The Qwen family
Alibaba's series, and the broadest size ladder available — Qwen3 runs from 0.6B to 32B dense, plus 30B-A3B and 235B-A22B as mixture-of-experts. The 152k vocabulary makes the embedding table a large share of a small model's file. Qwen2.5-Coder is the same architecture trained for code.