# Qwen3.8 27B: hardware requirements
> Qwen3.8 27B (Alibaba Qwen) is a 27B dense model with a 256K-token context window, released under the Apache 2.0 license. It needs about 18GB of memory at Q4_K_M with 8K context.
Qwen3.8 27B, released in August 2026, is one of the strongest open models that fits a single 24GB GPU, with image input, a 262K native context, and switchable thinking. It uses the same hybrid layout as the Qwen3.5 and 3.6 27B models, where three of every four layers use linear attention, so long contexts cost far less memory than on older dense models. At Q4_K_M it needs about 18GB with 8K context, leaving room for 64K or more of context on a 24GB card, and 32GB cards can run Q6_K. It supersedes Qwen3.6 27B and Qwen3 32B, but its long reasoning traces make speed on slower hardware a real consideration.

## Recommended hardware
- **Minimum GPU:** RTX 4060 Ti 16GB (16GB), Q3_K_M, ~17 tokens/sec
- **Best value:** RTX 3090 (24GB), Q4_K_M, ~39 tokens/sec
- **Fastest single GPU:** RTX 5090 (32GB), Q6_K, ~59 tokens/sec
- **Mac / unified memory:** Ryzen AI Max+ 395 (128GB), Q8_0, ~6.4 tokens/sec

## Memory by quantization
| Quant | Weights | Total @ 8K ctx | Total @ 32K ctx |
| --- | --- | --- | --- |
| Q2_K | 10.2GB | 11.6GB | 13.2GB |
| Q3_K_M | 13.2GB | 14.8GB | 16.4GB |
| Q4_K_M | 16.5GB | 18.1GB | 19.7GB |
| Q5_K_M | 19.8GB | 21.6GB | 23.2GB |
| Q6_K | 22.9GB | 24.8GB | 26.4GB |
| Q8_0 | 29.6GB | 31.6GB | 33.2GB |
| BF16 | 55.6GB | 58.4GB | 60.1GB |

## GPUs that run it fully in memory (fastest first)
| GPU | Memory | Quant | Speed | Price |
| --- | --- | --- | --- | --- |
| RTX 5090 | 32GB | Q6_K | ~59 tok/s | ~$4,300 |
| RTX 5080 | 16GB | Q3_K_M | ~55 tok/s | ~$1,600 |
| RTX 5070 Ti | 16GB | Q3_K_M | ~51 tok/s | ~$1,200 |
| RTX 4090 | 24GB | Q4_K_M | ~47 tok/s | ~$2,500 used |
| RTX PRO 6000 | 96GB | Q8_0 | ~46 tok/s | ~$16,000 |
| RTX 4080 Super | 16GB | Q3_K_M | ~43 tok/s | ~$1,150 used |
| RTX 3090 | 24GB | Q4_K_M | ~39 tok/s | ~$1,050 used |
| RTX 4070 Ti Super | 16GB | Q3_K_M | ~39 tok/s | ~$900 used |
| RX 7900 XTX | 24GB | Q4_K_M | ~36 tok/s | ~$1,050 used |
| RTX PRO 5000 72GB | 72GB | Q8_0 | ~35 tok/s | ~$9,100 |
| RX 9070 XT | 16GB | Q3_K_M | ~30 tok/s | ~$800 |
| RTX 5060 Ti 16GB | 16GB | Q3_K_M | ~26 tok/s | ~$620 |
| RTX 6000 Ada | 48GB | Q8_0 | ~25 tok/s | ~$4,750 used |
| V100 32GB | 32GB | Q6_K | ~25 tok/s | ~$700 used |
| M5 Ultra 512GB | 512GB | Q8_0 | ~22 tok/s | n/a |
| M5 Ultra 256GB | 256GB | Q8_0 | ~22 tok/s | ~$9,499 |
| MI50 32GB | 32GB | Q6_K | ~19 tok/s | ~$600 used |
| RTX A6000 | 48GB | Q8_0 | ~18 tok/s | ~$4,400 used |
| R9700 | 32GB | Q6_K | ~18 tok/s | ~$1,800 |
| RTX 4000 Ada | 20GB | Q4_K_M | ~17 tok/s | $1,250 MSRP |
| RTX 4060 Ti 16GB | 16GB | Q3_K_M | ~17 tok/s | ~$380 used |
| RX 9060 XT 16GB | 16GB | Q3_K_M | ~15 tok/s | ~$470 |
| M3 Ultra 512GB | 512GB | Q8_0 | ~15 tok/s | $9,499 MSRP |
| Arc Pro B60 | 24GB | Q4_K_M | ~15 tok/s | ~$650 |
| Arc Pro B70 | 32GB | Q6_K | ~14 tok/s | ~$1,300 |

## How to run it
- Ollama: `ollama run qwen3.8:27b`
- llama.cpp: `llama-server -hf unsloth/Qwen3.8-27B-GGUF:Q4_K_M -c 8192 -ngl 99`
- Weights: https://huggingface.co/Qwen/Qwen3.8-27B

## Strengths
- Top-tier reasoning and agentic coding for a model that fits a 24GB GPU
- Image input, toggleable thinking, and 262K native context
- Hybrid attention keeps the KV cache at 64 KiB per token
- Apache 2.0 license

## Weaknesses
- Uses many output tokens when thinking, so answers take a while
- Dense model: slower than 3B-active mixture-of-experts models on the same GPU
- Does not fit 16GB GPUs at Q4_K_M without offloading
Full page: https://llamabuilds.ai/models/qwen3-8-27b · Hardware guide: https://llamabuilds.ai/models/qwen3-8-27b/requirements

---
Source: https://llamabuilds.ai/models/qwen3-8-27b · Updated October 14, 2026 · Llama Builds. Speeds marked ~ are estimates from our bandwidth model (https://llamabuilds.ai/methodology); measured results are labeled as such.