Qwen2.5 VRAM requirements
Memory needed to run Qwen2.5 at every published size and quant, with an 8k context. Each figure is weights plus KV cache plus runtime buffers, from the same memory model as the VRAM calculator.
Total memory by size and quant
GB at an 8k context. Bold is the Q4_K_M column, the usual starting point.
| Size | Q8_0 | Q6_K | Q5_K_M | Q4_K_M | Q3_K_M |
|---|---|---|---|---|---|
| 0.5B | 2.8 | 2.7 | 2.6 | 2.6 | 2.5 |
| 1.5B · 0.9B active | 4.2 | 3.9 | 3.7 | 3.5 | 3.3 |
| 1.5B | 4.2 | 3.9 | 3.7 | 3.5 | 3.3 |
| 3B | 6.0 | 5.4 | 5.0 | 4.6 | 4.3 |
| 7B | 10.0 | 8.7 | 7.8 | 6.9 | 6.0 |
| 14B | 17 | 14 | 13 | 11 | 9.2 |
| 32B | 36 | 30 | 26 | 22 | 18 |
| 72B | 76 | 63 | 54 | 45 | 36 |
KV cache by context length
GB the context alone adds on top of the weights. Add this to a weights figure to size a longer session.
| Size | Weights at Q4_K_M | 4k ctx | 16k ctx | 32k ctx | 128k ctx |
|---|---|---|---|---|---|
| 0.5B | 0.3 | 0.3 | 1.1 | 2.1 | 8.6 |
| 1.5B · 0.9B active | 0.8 | 0.5 | 1.9 | 3.8 | 15 |
| 1.5B | 0.8 | 0.5 | 1.9 | 3.8 | 15 |
| 3B | 1.7 | 0.6 | 2.4 | 4.8 | 19 |
| 7B | 3.9 | 0.6 | 2.4 | 4.8 | 19 |
| 14B | 7.9 | 0.7 | 2.7 | 5.4 | 21 |
| 32B | 18 | 1.1 | 4.3 | 8.6 | 34 |
| 72B | 41 | 1.3 | 5.4 | 11 | 43 |
What each size fits on at Q4_K_M
Hardware whose memory holds the model with headroom, smallest pool first. Fits means under 85% of the pool at an 8k context. A Mac's pool is what macOS lets the GPU use - two thirds of its memory up to 32 GB, three quarters from 36 GB - not the whole of it.
- 0.5Bneeds about 2.6 GB
Smallest: AMD FirePro D700 6GB (6 GB), NVIDIA RTX 3050 6GB (6 GB), NVIDIA RTX 3060 Laptop 6GB (6 GB), NVIDIA RTX 4050 6GB (6 GB), and 131 more.
On a Mac: 8 GB or more, with macOS's default GPU limit.
- 1.5B · 0.9B activeneeds about 3.5 GB
Smallest: AMD FirePro D700 6GB (6 GB), NVIDIA RTX 3050 6GB (6 GB), NVIDIA RTX 3060 Laptop 6GB (6 GB), NVIDIA RTX 4050 6GB (6 GB), and 131 more.
On a Mac: 8 GB or more, with macOS's default GPU limit.
- 1.5Bneeds about 3.5 GB
Smallest: AMD FirePro D700 6GB (6 GB), NVIDIA RTX 3050 6GB (6 GB), NVIDIA RTX 3060 Laptop 6GB (6 GB), NVIDIA RTX 4050 6GB (6 GB), and 131 more.
On a Mac: 8 GB or more, with macOS's default GPU limit.
- 3Bneeds about 4.6 GB
Smallest: AMD FirePro D700 6GB (6 GB), NVIDIA RTX 3050 6GB (6 GB), NVIDIA RTX 3060 Laptop 6GB (6 GB), NVIDIA RTX 4050 6GB (6 GB), and 131 more.
On a Mac: 16 GB or more, with macOS's default GPU limit.
- 7Bneeds about 6.9 GB
Smallest: AMD RX 6700 (10 GB), NVIDIA P102-100 (10 GB), NVIDIA RTX 3080 10GB (10 GB), NVIDIA GTX 1080 Ti (11 GB), and 106 more.
On a Mac: 16 GB or more, with macOS's default GPU limit.
- 14Bneeds about 11 GB
Smallest: AMD MI50 16GB (16 GB), AMD RX 6800 16GB (16 GB), AMD RX 6800 XT 16GB (16 GB), AMD RX 6900 XT 16GB (16 GB), and 87 more.
On a Mac: 24 GB or more, with macOS's default GPU limit.
- 32Bneeds about 22 GB
Smallest: AMD MI50 32GB (32 GB), AMD Radeon AI PRO R9700 32GB (32 GB), AMD Radeon Pro V620 32GB (32 GB), AMD Radeon Pro W6800 (32 GB), and 41 more.
On a Mac: 36 GB or more, with macOS's default GPU limit.
- 72Bneeds about 45 GB
Smallest: AMD Instinct MI210 (64 GB), NVIDIA CMP 170HX 64GB (unlocked) (64 GB), NVIDIA RTX Pro 5000 Blackwell 72GB (72 GB), NVIDIA A100 80GB (80 GB), and 20 more.
On a Mac: 96 GB or more, with macOS's default GPU limit.
Frequently asked
How much VRAM does Qwen2.5 need?
From about 2.6 GB for the 0.5B model to about 45 GB for the 72B model, at Q4_K_M with an 8k context and including the KV cache and runtime buffers. Q8 needs more and Q3 less; the table above lists every size and rung.
Can I run Qwen2.5 on a 16GB GPU?
Yes. The 14B model needs about 11 GB at Q4_K_M with an 8k context, which fits a 16 GB card with headroom. Larger sizes need a smaller quant or a bigger card.
Can I run Qwen2.5 on a 24GB GPU such as an RTX 3090 or 4090?
Yes. The 14B model needs about 11 GB at Q4_K_M with an 8k context, which fits a 24 GB card with headroom. Larger sizes need a smaller quant or more memory.
How are these numbers calculated?
Weights at the quant's bits per weight, plus a KV cache sized from the model's attention design for the chosen context, plus compute buffers and a fixed runtime allowance. It is the same memory model the VRAM calculator uses. A family whose attention profile is not on file is sized as plain grouped-query attention, which errs towards needing more memory.