# Qwen3.5 9B: hardware requirements
> Qwen3.5 9B (Alibaba Qwen) is a 9B dense model with a 256K-token context window, released under the Apache 2.0 license. It needs about 7.2GB of memory at Q4_K_M with 8K context.
Qwen3.5 9B is the largest of the small Qwen3.5 models released in March 2026, with built-in image understanding, a 262K context, and switchable thinking. It uses Qwen's hybrid attention, where most layers are linear-attention layers with a fixed-size state, so long contexts cost far less memory than on older 8B models. It is one of the strongest models that fits an 8GB GPU: at Q4_K_M it needs about 7GB with 8K context, and 12GB cards can run Q8_0. It replaces Qwen3 8B and, for most uses, Qwen3 14B.

## Recommended hardware
- **Minimum GPU:** RTX 5060 (8GB), Q4_K_M, ~55 tokens/sec
- **Best value:** RTX 3060 12GB (12GB), Q6_K, ~30 tokens/sec
- **Fastest single GPU:** RTX 5090 (32GB), Q8_0, ~115 tokens/sec
- **Mac / unified memory:** Ryzen AI Max+ 395 (128GB), Q8_0, ~17 tokens/sec

## Memory by quantization
| Quant | Weights | Total @ 8K ctx | Total @ 32K ctx |
| --- | --- | --- | --- |
| Q2_K | 3.8GB | 4.8GB | 5.6GB |
| Q3_K_M | 5.0GB | 6.0GB | 6.8GB |
| Q4_K_M | 6.2GB | 7.2GB | 8.0GB |
| Q5_K_M | 7.3GB | 8.3GB | 9.1GB |
| Q6_K | 8.4GB | 9.5GB | 10.3GB |
| Q8_0 | 10.8GB | 12.0GB | 12.8GB |
| BF16 | 20.4GB | 21.8GB | 22.6GB |

## GPUs that run it fully in memory (fastest first)
| GPU | Memory | Quant | Speed | Price |
| --- | --- | --- | --- | --- |
| RTX 5090 | 32GB | Q8_0 | ~115 tok/s | ~$4,300 |
| RTX PRO 6000 | 96GB | Q8_0 | ~115 tok/s | ~$16,000 |
| RTX PRO 5000 72GB | 72GB | Q8_0 | ~90 tok/s | ~$9,100 |
| RTX 4090 | 24GB | Q8_0 | ~69 tok/s | ~$2,500 used |
| RTX 5080 | 16GB | Q8_0 | ~66 tok/s | ~$1,600 |
| RTX 6000 Ada | 48GB | Q8_0 | ~66 tok/s | ~$4,750 used |
| RTX 5070 Ti | 16GB | Q8_0 | ~62 tok/s | ~$1,200 |
| RTX 5070 | 12GB | Q6_K | ~60 tok/s | ~$800 |
| RTX 3090 | 24GB | Q8_0 | ~59 tok/s | ~$1,050 used |
| M5 Ultra 512GB | 512GB | Q8_0 | ~57 tok/s | n/a |
| M5 Ultra 256GB | 256GB | Q8_0 | ~57 tok/s | ~$9,499 |
| RTX 5060 | 8GB | Q4_K_M | ~55 tok/s | ~$450 |
| RX 7900 XTX | 24GB | Q8_0 | ~53 tok/s | ~$1,050 used |
| RTX 4080 Super | 16GB | Q8_0 | ~52 tok/s | ~$1,150 used |
| V100 32GB | 32GB | Q8_0 | ~51 tok/s | ~$700 used |
| RTX A6000 | 48GB | Q8_0 | ~49 tok/s | ~$4,400 used |
| RTX 4070 Ti Super | 16GB | Q8_0 | ~47 tok/s | ~$900 used |
| RTX 4070 | 12GB | Q6_K | ~46 tok/s | ~$560 used |
| MI50 32GB | 32GB | Q8_0 | ~39 tok/s | ~$600 used |
| M3 Ultra 512GB | 512GB | Q8_0 | ~39 tok/s | $9,499 MSRP |
| R9700 | 32GB | Q8_0 | ~36 tok/s | ~$1,800 |
| RX 9070 XT | 16GB | Q8_0 | ~36 tok/s | ~$800 |
| M5 Max 128GB | 128GB | Q8_0 | ~35 tok/s | ~$5,100 |
| RTX 5060 Ti 16GB | 16GB | Q8_0 | ~32 tok/s | ~$620 |
| M4 Max 128GB | 128GB | Q8_0 | ~31 tok/s | $3,499 MSRP |

## How to run it
- Ollama: `ollama run qwen3.5:9b`
- llama.cpp: `llama-server -hf unsloth/Qwen3.5-9B-GGUF:Q4_K_M -c 8192 -ngl 99`
- Weights: https://huggingface.co/Qwen/Qwen3.5-9B

## Strengths
- Image input, toggleable thinking, and 262K context in an 8GB-class model
- Hybrid attention keeps long-context memory low
- Strong reasoning and coding for its size
- Apache 2.0 license

## Weaknesses
- Thinking mode is verbose, so answers take longer on slow hardware
- Limited world knowledge, as with any 9B model
- Linear-attention layers need recent runtime versions for full speed
Full page: https://llamabuilds.ai/models/qwen3-5-9b · Hardware guide: https://llamabuilds.ai/models/qwen3-5-9b/requirements

---
Source: https://llamabuilds.ai/models/qwen3-5-9b · Updated October 14, 2026 · Llama Builds. Speeds marked ~ are estimates from our bandwidth model (https://llamabuilds.ai/methodology); measured results are labeled as such.