What a Qwen Image 2.1 GGUF is
GGUF is a single-file container for quantized model weights. For Qwen-Image-2.1 it holds the denoising transformer at reduced precision. It does not include the text encoder or the VAE, so those are downloaded separately.
The Qwen team publishes full-precision weights; GGUF conversions are made by community members such as unsloth, leejet and Abiray on Hugging Face. They are derivatives of the original model and their repositories declare the same Qwen Research License.
GGUF quantization sizes compared
File sizes from the unsloth/Qwen-Image-2.1-GGUF repository, read on 4 October 2026. Other repositories differ by a few hundred megabytes for the same level because they keep different layers at higher precision.
| Quant level | File size | Notes |
|---|---|---|
| Q2_K | 2.30 GiB | Smallest file; fine detail and lettering may suffer |
| Q3_K_S / Q3_K_M | 2.54 / 2.95 GiB | Very low memory budgets |
| Q4_K_S / Q4_K_M | 3.64 / 3.91 GiB | Common starting point on mid-range cards |
| Q5_K_S / Q5_K_M | 4.19 / 5.02 GiB | Middle ground |
| Q6_K | 5.84 GiB | Popular high-precision choice |
| Q8_0 | 7.12 GiB | Nearest to full precision among the quantized files |
| F16 | 13.25 GiB | Unquantized 16-bit, same size class as the BF16 original |
The same Q6_K level is 5.59 GiB in leejet/Qwen-Image-2.1-GGUF and 5.47 GiB in Abiray/Qwen-Image-2.1-GGUF. Whichever repository you choose, add the text encoder to your memory budget: the INT8 Qwen3-VL 8B encoder in the ComfyUI repack is about 9.35 GB and the W4A8 variant about 6.3 GB.
Which quant to choose: Q4_K_M, Q6_K or Q8_0
There is no official quality table for these files, so treat the following as community guidance rather than benchmarks. Higher numbers keep more bits per weight and stay closer to the original model; lower numbers trade fidelity for memory.
- Q6_K is the most searched level for a reason: it is about 1.3 GiB smaller than Q8_0 while staying in the high-precision group. One quantizer's model card calls it the high-quality sweet spot and suggests 10 to 12 GB of VRAM.
- Q4_K_M is the usual choice when VRAM is limited; the same card suggests 6 to 8 GB.
- Q8_0 is the safest pick if it fits, because it changes the weights least.
- Q2 and Q3 files load on very small cards, but community guides warn that fine detail, anatomy, typography and prompt adherence can suffer.
If you have the memory for it, the INT8 file from Comfy-Org (about 7.3 GB) runs in the stock ComfyUI templates with no custom node at all.
Set up a Qwen Image 2.1 GGUF in ComfyUI
- Update ComfyUI. Qwen-Image-2.1 support relies on recent core nodes; the documentation notes that missing nodes usually mean an outdated build.
- Install the ComfyUI-GGUF custom node by city96 through ComfyUI Manager, then restart.
- Download one .gguf file, for example the Q6_K, and place it in ComfyUI/models/unet (ComfyUI/models/diffusion_models also works).
- Download a Qwen3-VL 8B text encoder from Comfy-Org/Qwen-Image-2.1, such as qwen3vl_8b_int8_convrot.safetensors, into ComfyUI/models/text_encoders.
- Download qwen_image_2.1_vae_bf16.safetensors into ComfyUI/models/vae.
- Open a GGUF workflow (several GGUF repositories include one), or take the official Qwen-Image-2.1 text-to-image template and swap its diffusion model loader for the Unet Loader (GGUF) node pointing at your file.
- Queue a prompt. Start with the template defaults before changing anything.
| Setting | Value | Note |
|---|---|---|
| Steps | 25 | The published pipeline uses about 40 to 50; raise steps only if detail fails to resolve |
| CFG | 1 | Negative prompts have no effect at CFG 1; CFG 2 follows dense prompts more closely |
| Sampler / scheduler | euler / simple | Same for all three official templates |
| Resolution | 1.0 MP is about 1024 x 1024 | Set about 4.0 MP for native 2048 x 2048 |
Running a GGUF without ComfyUI
GGUF files are not tied to ComfyUI. The unsloth model card documents running them with stable-diffusion.cpp, passing the GGUF as the diffusion model together with the VAE and a GGUF build of the Qwen3-VL 8B text encoder.
sd-cli --diffusion-model qwen-image-2.1-Q4_K_M.gguf \
--vae qwen_image_2.1_vae_bf16.safetensors \
--llm Qwen3-VL-8B-Instruct-UD-Q4_K_XL.gguf \
-p "a cartoon sloth mascot waving, flat vector illustration" \
--steps 20 --cfg-scale 6.0 --sampling-method euler -W 1024 -H 1024 \
-o out.pngBefore you install anything
A working GGUF setup still means 10 GB or more of downloads, a custom node and some troubleshooting. If you only want to see how Qwen Image 2.1 responds to your prompt, generate in this studio first. This studio uses a hosted API rather than your hardware, at 1 credit per image.
qwenimage21.pro is an independently operated workspace. It is not affiliated with or endorsed by Alibaba or the Qwen team, and it does not host or redistribute model weights.