# Qwen3 32B: hardware requirements
> Qwen3 32B (Alibaba Qwen) is a 32.8B dense model with a 32K-token context window, released under the Apache 2.0 license. It needs about 23GB of memory at Q4_K_M with 8K context.
Qwen3 32B was the strongest dense model most people could run on a single 24GB GPU in 2025, and it is our reference point for 32B-class speed. It has been superseded by the Qwen3.5, 3.6, and 3.8 27B models, which are smaller, add vision and a 262K context, and score far higher on reasoning and agentic coding. At Q4_K_M it needs about 23GB with 8K context, so a 24GB card fits it with little room for longer context; 32GB cards and Macs with 36GB or more are more comfortable. Because it is dense, every token reads all 32B parameters, so it is much slower than a 30B-A3B mixture-of-experts model on the same GPU.

## Recommended hardware
- **Minimum GPU:** RTX 4000 Ada (20GB), Q3_K_M, ~18 tokens/sec
- **Best value:** RTX 3090 (24GB), Q3_K_M, ~41 tokens/sec
- **Fastest single GPU:** RTX 5090 (32GB), Q5_K_M, ~58 tokens/sec
- **Mac / unified memory:** Ryzen AI Max+ 395 (128GB), Q8_0, ~5.5 tokens/sec

## Memory by quantization
| Quant | Weights | Total @ 8K ctx | Total @ 32K ctx |
| --- | --- | --- | --- |
| Q2_K | 12.2GB | 15.4GB | 21.8GB |
| Q3_K_M | 15.9GB | 19.1GB | 25.6GB |
| Q4_K_M | 19.8GB | 23.1GB | 29.6GB |
| Q5_K_M | 23.3GB | 26.7GB | 33.2GB |
| Q6_K | 26.9GB | 30.5GB | 36.9GB |
| Q8_0 | 34.8GB | 38.6GB | 45.0GB |
| BF16 | 65.5GB | 70.2GB | 76.7GB |

## GPUs that run it fully in memory (fastest first)
| GPU | Memory | Quant | Speed | Price |
| --- | --- | --- | --- | --- |
| RTX 5090 | 32GB | Q5_K_M | ~58 tok/s | ~$4,300 |
| RTX 4090 | 24GB | Q3_K_M | ~48 tok/s | ~$2,500 used |
| RTX 3090 | 24GB | Q3_K_M | ~41 tok/s | ~$1,050 used |
| RTX PRO 6000 | 96GB | Q8_0 | ~40 tok/s | ~$16,000 |
| RX 7900 XTX | 24GB | Q3_K_M | ~37 tok/s | ~$1,050 used |
| RTX PRO 5000 72GB | 72GB | Q8_0 | ~30 tok/s | ~$9,100 |
| V100 32GB | 32GB | Q5_K_M | ~24 tok/s | ~$700 used |
| RTX 6000 Ada | 48GB | Q8_0 | ~22 tok/s | ~$4,750 used |
| MI50 32GB | 32GB | Q5_K_M | ~19 tok/s | ~$600 used |
| M5 Ultra 512GB | 512GB | Q8_0 | ~19 tok/s | n/a |
| M5 Ultra 256GB | 256GB | Q8_0 | ~19 tok/s | ~$9,499 |
| RTX 4000 Ada | 20GB | Q3_K_M | ~18 tok/s | $1,250 MSRP |
| R9700 | 32GB | Q5_K_M | ~17 tok/s | ~$1,800 |
| RTX A6000 | 48GB | Q8_0 | ~16 tok/s | ~$4,400 used |
| Arc Pro B60 | 24GB | Q3_K_M | ~15 tok/s | ~$650 |
| Arc Pro B70 | 32GB | Q5_K_M | ~14 tok/s | ~$1,300 |
| M3 Ultra 512GB | 512GB | Q8_0 | ~13 tok/s | $9,499 MSRP |
| Tesla P40 | 24GB | Q3_K_M | ~12 tok/s | ~$230 used |
| M5 Max 128GB | 128GB | Q8_0 | ~11 tok/s | ~$5,100 |
| M4 Max 128GB | 128GB | Q8_0 | ~10 tok/s | $3,499 MSRP |
| DGX Spark | 128GB | Q8_0 | ~5.8 tok/s | ~$6,950 |
| Ryzen AI Max+ PRO 495 | 192GB | Q8_0 | ~5.8 tok/s | ~$6,800 |
| Ryzen AI Max+ 395 | 128GB | Q8_0 | ~5.5 tok/s | ~$3,300 |

## How to run it
- Ollama: `ollama run qwen3:32b`
- llama.cpp: `llama-server -hf Qwen/Qwen3-32B-GGUF:Q4_K_M -c 8192 -ngl 99`
- MLX (Mac): `mlx_lm.chat --model mlx-community/Qwen3-32B-4bit`
- Weights: https://huggingface.co/Qwen/Qwen3-32B

## Strengths
- Toggleable thinking with strong math and coding for a 2025 model
- Dense weights quantize predictably; Q4_K_M loses little quality
- Apache 2.0 license and broad runtime support

## Weaknesses
- Superseded by Qwen3.8 27B, which is smaller and much stronger
- Large KV cache (256 KiB per token) limits context on 24GB cards
- Native context is only 32K tokens
- Text only
Full page: https://llamabuilds.ai/models/qwen3-32b · Hardware guide: https://llamabuilds.ai/models/qwen3-32b/requirements

---
Source: https://llamabuilds.ai/models/qwen3-32b · Updated October 14, 2026 · Llama Builds. Speeds marked ~ are estimates from our bandwidth model (https://llamabuilds.ai/methodology); measured results are labeled as such.