Qwen3 32B (Q6_K_XL) — 19.5 GBon NVIDIA RTX 4000 SFF

Q6_K_XL NVIDIA RTX 4000 SFF

Overview

Qwen3 32B is a 32B parameter dense language model by Qwen, with code, multilingual, thinking, tool-calls capabilities. It supports a context window of up to 40,960 tokens.

Qwen3 32B is a 32-billion-parameter dense transformer from Alibaba's Qwen team, combining thinking capabilities with strong code generation, tool calling, and multilingual support. It occupies a mid-range parameter class that balances reasoning depth with practical deployment requirements, outperforming many larger models on math and logic benchmarks. The model supports 14 languages including English, Chinese, and Arabic. With a 40K context window and flash attention, it fits on a single high-end GPU at Q4 quantization for self-hosted inference.

At Q6_K_XL quantization (high quality tier), the model weighs 26.97 GB. This exceeds the 20 GB of VRAM on NVIDIA RTX 4000 SFF. Inference is still possible via CPU offload or memory-mapped loading from disk, but expect significantly reduced performance.

Hardware Requirements

Model size	26.97 GB
VRAM available	20 GB
VRAM used	19.5 GB
Min RAM required	8.8 GB
GPU layers	43 / 64
Context size	512
Backend	cuda13
Flash attention	Yes

Performance Notes

Reduced performance

21 of 64 layers run on CPU. Your system needs at least 8.8 GB RAM. If system RAM is insufficient, data will be read from disk on demand, with significant performance impact.

Shorter conversations

Context size has been reduced due to memory constraints. This may affect the model's ability to handle long conversations.

Deploy

Server Local

Prerequisites

Ensure your GPU nodes are prepared with the NVIDIA container toolkit:

ansible-playbook prositronic.infra.nvidia_container_toolkit

Command

helmfile --state-values-file <(curl -s https://www.prositronic.eu/values/qwen3-32b/q6_k_xl/nvidia-rtx4000sff.yaml) apply

Generated values.yaml

/values/qwen3-32b/q6_k_xl/nvidia-rtx4000sff.yaml

Loading values…

Install llama.cpp

Install llama.cpp from the official build instructions for your platform and backend.

Download Model

curl -L -o qwen3-32b.gguf "https://huggingface.co/unsloth/Qwen3-32B-GGUF/resolve/main/Qwen3-32B-UD-Q6_K_XL.gguf"

Start Server

llama-server \
  -m qwen3-32b.gguf \
  --n-gpu-layers 43 \
  --ctx-size 512 \
  --flash-attn

Verify

curl http://localhost:8080/health

Explore

• Configure a deployment
• Browse models
• Browse hardware

Other quantizations

• Qwen3 32B (Q8_0)
• Qwen3 32B (Q5_K_M)
• Qwen3 32B (Q4_K_M)

Other hardware

• Qwen3 32B on CPU Only
• Qwen3 32B on NVIDIA A100 80GB
• Qwen3 32B on NVIDIA H100

Frequently Asked Questions

How much VRAM does Qwen3 32B (Q6_K_XL) need?

The Q6_K_XL quantization of Qwen3 32B requires 26.97 GB. 43 of 64 layers fit in the 20 GB of VRAM on NVIDIA RTX 4000 SFF; remaining layers are offloaded to CPU.

Can I run Qwen3 32B on NVIDIA RTX 4000 SFF?

Yes, with reduced performance. NVIDIA RTX 4000 SFF can run Qwen3 32B (Q6_K_XL), but only 43 of 64 layers fit in VRAM. The rest are offloaded to CPU.

What is quantization?

Quantization reduces a model's numerical precision from its original floating-point format to a more compact representation. This shrinks the file size and VRAM footprint, making it possible to run large models on consumer hardware. The trade-off is a small reduction in output quality. Q6_K_XL compresses Qwen3 32B from its original size down to 26.97 GB.

What quantization should I choose for Qwen3 32B?

Q6_K_XL is a high-quality quantization. Higher-quality quants (Q8, Q6) preserve more model accuracy but need more VRAM. Lower quants (Q4, Q3, Q2) reduce VRAM usage at the cost of some quality. Choose based on your available hardware and quality requirements.

What is flash attention and why is it enabled?

Flash attention is a memory-efficient algorithm that speeds up the attention mechanism in transformer models. It reduces VRAM usage during inference by avoiding the materialisation of the full attention matrix. For Qwen3 32B on NVIDIA RTX 4000 SFF, flash attention is enabled to maximise context length and throughput within the available 20 GB of VRAM.

Why are some layers offloaded to CPU?

NVIDIA RTX 4000 SFF has 20 GB of VRAM, but Qwen3 32B (Q6_K_XL) requires approximately 26.97 GB. Only 43 of 64 layers fit in VRAM; the remaining layers run on CPU, which is slower but still functional.

How do I run Qwen3 32B (Q6_K_XL) with Ollama?

Run ollama run qwen3:32b-q6_k_xl to start Qwen3 32B (Q6_K_XL). Ollama handles downloading the model weights automatically on first run.

Last updated: March 5, 2026