# Qwen3 8B: hardware requirements
> Qwen3 8B (Alibaba Qwen) is a 8.2B dense model with a 32K-token context window, released under the Apache 2.0 license. It needs about 7GB of memory at Q4_K_M with 8K context.
Qwen3 8B brought switchable thinking to the 8B class in April 2025 and was one of the most capable models you could run on an 8GB GPU. It has since been superseded by Qwen3.5 9B, which adds vision, a 262K context, and much stronger reasoning at a similar size. At Q4_K_M it needs about 7GB with 8K context, so it fits 8GB cards, and 12GB cards can run Q8_0. The native context is 32K, extendable to 128K with YaRN at the cost of a large KV cache.

## Recommended hardware
- **Minimum GPU:** RTX 5060 (8GB), Q4_K_M, ~67 tokens/sec
- **Best value:** RTX 5060 (8GB), Q4_K_M, ~67 tokens/sec
- **Fastest single GPU:** RTX 5090 (32GB), Q8_0, ~140 tokens/sec
- **Mac / unified memory:** Ryzen AI Max+ 395 (128GB), Q8_0, ~21 tokens/sec

## Memory by quantization
| Quant | Weights | Total @ 8K ctx | Total @ 32K ctx |
| --- | --- | --- | --- |
| Q2_K | 3.1GB | 5.0GB | 8.6GB |
| Q3_K_M | 4.0GB | 6.0GB | 9.6GB |
| Q4_K_M | 5.0GB | 7.0GB | 10.6GB |
| Q5_K_M | 5.9GB | 7.9GB | 11.5GB |
| Q6_K | 6.8GB | 8.9GB | 12.5GB |
| Q8_0 | 8.7GB | 10.8GB | 14.4GB |
| BF16 | 16.4GB | 18.7GB | 22.3GB |

## GPUs that run it fully in memory (fastest first)
| GPU | Memory | Quant | Speed | Price |
| --- | --- | --- | --- | --- |
| RTX 5090 | 32GB | Q8_0 | ~140 tok/s | ~$4,300 |
| RTX PRO 6000 | 96GB | Q8_0 | ~140 tok/s | ~$16,000 |
| RTX PRO 5000 72GB | 72GB | Q8_0 | ~110 tok/s | ~$9,100 |
| RTX 4090 | 24GB | Q8_0 | ~85 tok/s | ~$2,500 used |
| RTX 5080 | 16GB | Q8_0 | ~81 tok/s | ~$1,600 |
| RTX 6000 Ada | 48GB | Q8_0 | ~81 tok/s | ~$4,750 used |
| RTX 5070 Ti | 16GB | Q8_0 | ~76 tok/s | ~$1,200 |
| RTX 5070 | 12GB | Q6_K | ~73 tok/s | ~$800 |
| RTX 3090 | 24GB | Q8_0 | ~72 tok/s | ~$1,050 used |
| M5 Ultra 512GB | 512GB | Q8_0 | ~69 tok/s | n/a |
| M5 Ultra 256GB | 256GB | Q8_0 | ~69 tok/s | ~$9,499 |
| RTX 5060 | 8GB | Q4_K_M | ~67 tok/s | ~$450 |
| RX 7900 XTX | 24GB | Q8_0 | ~65 tok/s | ~$1,050 used |
| RTX 4080 Super | 16GB | Q8_0 | ~63 tok/s | ~$1,150 used |
| V100 32GB | 32GB | Q8_0 | ~62 tok/s | ~$700 used |
| RTX A6000 | 48GB | Q8_0 | ~60 tok/s | ~$4,400 used |
| RTX 4070 Ti Super | 16GB | Q8_0 | ~58 tok/s | ~$900 used |
| RTX 4070 | 12GB | Q6_K | ~56 tok/s | ~$560 used |
| M3 Ultra 512GB | 512GB | Q8_0 | ~48 tok/s | $9,499 MSRP |
| MI50 32GB | 32GB | Q8_0 | ~48 tok/s | ~$600 used |
| R9700 | 32GB | Q8_0 | ~45 tok/s | ~$1,800 |
| RX 9070 XT | 16GB | Q8_0 | ~45 tok/s | ~$800 |
| M5 Max 128GB | 128GB | Q8_0 | ~43 tok/s | ~$5,100 |
| RTX 5060 Ti 16GB | 16GB | Q8_0 | ~40 tok/s | ~$620 |
| M4 Max 128GB | 128GB | Q8_0 | ~38 tok/s | $3,499 MSRP |

## How to run it
- Ollama: `ollama run qwen3:8b`
- llama.cpp: `llama-server -hf Qwen/Qwen3-8B-GGUF:Q4_K_M -c 8192 -ngl 99`
- MLX (Mac): `mlx_lm.chat --model mlx-community/Qwen3-8B-4bit`
- Weights: https://huggingface.co/Qwen/Qwen3-8B

## Strengths
- Thinking mode can be switched on for hard problems and off for quick answers
- Fits 8GB GPUs at Q4_K_M with 8K context
- Broad multilingual coverage (119 languages and dialects) and solid tool calling
- Apache 2.0 license with no usage restrictions

## Weaknesses
- Superseded by Qwen3.5 9B, which is better on nearly every task
- Native context is only 32K tokens
- Thinking mode is verbose, which slows answers on low-bandwidth hardware
- Limited world knowledge, typical of 8B models
Full page: https://llamabuilds.ai/models/qwen3-8b · Hardware guide: https://llamabuilds.ai/models/qwen3-8b/requirements

---
Source: https://llamabuilds.ai/models/qwen3-8b · Updated October 14, 2026 · Llama Builds. Speeds marked ~ are estimates from our bandwidth model (https://llamabuilds.ai/methodology); measured results are labeled as such.