Qwen 2.5 32B VRAM Requirements
Qwen 2.5 32B needs about 21.4 GB of VRAM at Q4_K_M with an 8K context: 18.5 GB for weights, 2.0 GB for the KV cache and 0.9 GB of runtime overhead. At Q8_0 it needs 35.3 GB and at full FP16 precision 64.0 GB. The smallest GPU that runs it fully is the NVIDIA L4 24GB.
21.4 GB VRAM
Qwen 2.5 32B at Q4_K_M with 8K context and the KV cache in FP16.
- Weights 18.5 GB
- KV cache 2.0 GB
- Runtime overhead 0.9 GB
Which GPUs can run it
- RTX 3060 12GB53% on GPU~3.5 tok/s
- RTX 4060 Ti 16GB70% on GPU~4.4 tok/s
- RTX 3090 24GBFits~28 tok/s
- Radeon RX 7900 XTX 24GBFits~25 tok/s
- RTX 4090 24GBFits~31 tok/s
- RTX 5090 32GBFits~54 tok/s
- RTX 6000 Ada 48GBFits~29 tok/s
- A100 80GBFits~61 tok/s
- H100 80GB SXMFits~97 tok/s
- Ryzen AI Max+ 395 128GBFits~7.3 tok/s
- Apple M4 Max 128GBFits~17 tok/s
- Apple M5 Max 128GBFits~19 tok/s
- RTX PRO 6000 Blackwell 96GBFits~54 tok/s
- NVIDIA DGX Spark 128GBFits~7.8 tok/s
- Apple M3 Ultra 512GBFits~25 tok/s
Qwen 2.5 32B VRAM by quantization
Total memory with an FP16 KV cache. "Smallest tier" is the smallest standard VRAM size with a 6% safety margin at 8K.
| Quantization | Bits/weight | Weights | Total at 8K | Total at 32K | Smallest tier |
|---|---|---|---|---|---|
| Q4_K_M | 4.85 | 18.5 GB | 21.4 GB | 27.4 GB | 24 GB |
| Q5_K_M | 5.69 | 21.7 GB | 24.6 GB | 30.6 GB | 32 GB |
| Q6_K | 6.59 | 25.2 GB | 28.0 GB | 34.0 GB | 32 GB |
| Q8_0 | 8.50 | 32.5 GB | 35.3 GB | 41.3 GB | 48 GB |
| FP16 / BF16 | 16.00 | 61.1 GB | 64.0 GB | 70.0 GB | 80 GB |
| Q3_K_M | 3.91 | 14.9 GB | 17.8 GB | 23.8 GB | 20 GB |
| Q2_K | 3.00 | 11.5 GB | 14.3 GB | 20.3 GB | 16 GB |
VRAM by context length
At Q4_K_M. The KV cache grows linearly with context: 0.25 GB per 1,000 tokens in FP16.
| Context | KV cache (FP16) | KV cache (Q8_0) | Total, FP16 cache | Total, Q8_0 cache |
|---|---|---|---|---|
| 4K tokens | 1.0 GB | 0.5 GB | 20.4 GB | 19.9 GB |
| 8K tokens | 2.0 GB | 1.1 GB | 21.4 GB | 20.5 GB |
| 32K tokens | 8.0 GB | 4.3 GB | 27.4 GB | 23.6 GB |
GPUs that can run Qwen 2.5 32B
At Q4_K_M with 8K context. "Best quant" is the highest-quality format that fits fully; "max context" is at Q4_K_M.
| GPU | Usable memory | Status | Best quant | Max context | Speed |
|---|---|---|---|---|---|
| NVIDIA L4 24GB | 24.0 GB | Fits | Q4_K_M | 12K | ~9.2 tok/s |
| Tesla P40 24GB | 24.0 GB | Fits | Q4_K_M | 12K | ~9.1 tok/s |
| Apple M4 Max 36GB | 24.0 GB | Fits | Q4_K_M | 12K | ~13 tok/s |
| Intel Arc Pro B60 24GB | 24.0 GB | Fits | Q4_K_M | 12K | ~11 tok/s |
| RTX 3090 24GB | 24.0 GB | Fits | Q4_K_M | 12K | ~28 tok/s |
| Radeon RX 7900 XTX 24GB | 24.0 GB | Fits | Q4_K_M | 12K | ~25 tok/s |
| RTX 3090 Ti 24GB | 24.0 GB | Fits | Q4_K_M | 12K | ~31 tok/s |
| RTX 4090 24GB | 24.0 GB | Fits | Q4_K_M | 12K | ~31 tok/s |
| Radeon AI PRO R9700 32GB | 32.0 GB | Fits | Q6_K | 32K | ~17 tok/s |
| RTX 5090 32GB | 32.0 GB | Fits | Q6_K | 32K | ~54 tok/s |
| Apple M4 Pro 48GB | 36.0 GB | Fits | Q6_K | 32K | ~8.4 tok/s |
| A100 40GB | 40.0 GB | Fits | Q8_0 | 32K | ~47 tok/s |
| Apple M5 Pro 64GB | 48.0 GB | Fits | Q8_0 | 32K | ~9.5 tok/s |
| Apple M1 Max 64GB | 48.0 GB | Fits | Q8_0 | 32K | ~12 tok/s |
| Radeon RX 7900 XT 20GB | 20.0 GB | 88% on GPU | Q3_K_M | – | ~9.7 tok/s |
| RTX 4060 Ti 16GB | 16.0 GB | 70% on GPU | Q2_K | – | ~4.4 tok/s |
| RTX 4070 Ti Super 16GB | 16.0 GB | 70% on GPU | Q2_K | – | ~5.5 tok/s |
| RTX 4080 16GB | 16.0 GB | 70% on GPU | Q2_K | – | ~5.5 tok/s |
| RTX 4080 Super 16GB | 16.0 GB | 70% on GPU | Q2_K | – | ~5.6 tok/s |
| RTX 5060 Ti 16GB | 16.0 GB | 70% on GPU | Q2_K | – | ~5.0 tok/s |
About Qwen 2.5 32B
The model that made 24 GB cards the standard for local AI. Q4_K_M weights are about 18.5 GB, which fits on an RTX 3090 or 4090 with roughly 12K of context. The same architecture underlies QwQ-32B and DeepSeek R1 Distill Qwen 32B.
| Spec | Value |
|---|---|
| Parameters | 32.8B |
| Layers | 64 |
| Hidden size | 5120 |
| Attention heads / KV heads | 40 / 8 |
| Head dimension | 128 |
| Vocabulary | 152,064 |
| Max context (used here) | 32K tokens |
| KV cache per token (FP16) | 256 KB |
| Released | 2024-09 |
| License | Apache 2.0 |
Qwen 2.5 32B VRAM: common questions
Can Qwen 2.5 32B run on 16GB VRAM?
Only at Q3_K_M or lower with a short context, or with partial CPU offload. 24 GB is the comfortable minimum for Q4_K_M.
How much VRAM does Qwen 2.5 32B need?
About 21.4 GB at Q4_K_M with an 8K context, 35.3 GB at Q8_0 and 64.0 GB at FP16. Longer contexts add 0.25 GB per 1,000 tokens with an FP16 cache.
What is the cheapest GPU that can run Qwen 2.5 32B?
The smallest listed device that fits it fully at Q4_K_M with 8K context is the NVIDIA L4 24GB (24.0 GB usable).