Qwen/Qwen-Image-2.1 7B

• License: Qwen Research License Agreement
• Release date: 2026-09-20 (official release announcement)
• Modalities: Input: text, image, multiple images; Output: image
• Original model: Qwen/Qwen-Image-2.1

Description

Qwen-Image-2.1 unifies text-to-image generation and reference-image editing, including native transparent images, subject extraction and transparent-layer editing. Its 7B visual generation component uses 32 Single-Stream DiT layers; the parameter count does not include the separate text encoder. The model focuses on typography, realistic textures and preserving people and products across edits.

Run

self serve qwen-image-2.1:7b

Here is a link to the API docs: Image API.

Example Request

With the server running, open another terminal and send a request:

curl http://localhost:8080/v1/images/generations \
	-H 'Content-Type: application/json' \
	-d '{
		"prompt": "A red fox in a snowy forest",
		"size": "1024x1024",
		"n": 1,
		"output_format": "png"
	}' \
	-o response.json

The server returns JSON with base64-encoded image bytes in data[0].b64_json, not a hosted image URL. Decode the first result with jq and base64:

jq -r '.data[0].b64_json' response.json | base64 --decode > output.png

Engine Parameters

Served by the unmodified upstream stable-diffusion.cpp sd-server through the sd-server adapter, with one loaded pipeline per process. These are startup settings; defaults below come from the registry. Unset settings are delegated to the engine rather than assigned a default by the adapter.

Parameter	Default	What it controls
steps	40	Sampling steps; accepts integers from 1 to 200.
cfg_scale	6.0	Guidance scale; accepts values from 0 to 50.
sampling_method	euler	Sampling method; the adapter accepts only euler.
flash_attention	true	Enable flash attention.
offload_to_cpu	true	Offload inactive model stages to CPU.
threads	Unset (engine default)	CPU workers; accepts integers from 1 to 1024.
seed	Unset (engine default)	Default seed; accepts integers from 0 to 9223372036854775807.

Override startup settings with --set <parameter>=<value>.

Per-request fields such as size, n and output_format belong in the API request, not --set; a request's seed overrides the startup seed. Backend sampling overrides can be supplied through sd_cpp_extra_args.sample_params; see the Image API for the backend-specific request format. Omitting overrides keeps the startup sampling defaults.

Quantizations and Files

Default variant: int8. Sizes and SHA-256 hashes are pinned in the model registry. Each variant requires all four files listed, including the vision projector.

int8 (default)

File	Role	Size	SHA-256	Origin
diffusion_models/qwen_image_2.1_int8_convrot.safetensors	model	6.76 GiB (7256783064 bytes)	cb74113cb03faecd79611b01fd7fd642f0aa60d6f0b95086abee214d75eaa57d	Comfy-Org/Qwen-Image-2.1
vae/qwen_image_2.1_vae_bf16.safetensors	vae	644.2 MiB (675509688 bytes)	bb21f7473051e1ac368515dd3f2e15cd44d7a11748ee8823e1ddca3e4876b7c9	Comfy-Org/Qwen-Image-2.1
Qwen3VL-8B-Instruct-Q4_K_M.gguf	text-encoder	4.68 GiB (5027784800 bytes)	67d1659bfe71b89d50b45a4ad1a9e5b997e5bb16ce5da66a6a6167abd569e9e2	Qwen/Qwen3-VL-8B-Instruct-GGUF
mmproj-Qwen3VL-8B-Instruct-F16.gguf	mmproj	1.08 GiB (1159029824 bytes)	ca524100ebf825c9a870db1c580d03879e0da0ab2541697e2458e64891cf9d38	Qwen/Qwen3-VL-8B-Instruct-GGUF

Total download: 13.15 GiB (14119107376 bytes).

q4

File	Role	Size	SHA-256	Origin
qwen-image-2.1-Q4_K_M.gguf	model	3.91 GiB (4199565024 bytes)	631d532e7ca71e8d90a87c71d3699761a812039d22e3370e87498d87754660fe	unsloth/Qwen-Image-2.1-GGUF
vae/qwen_image_2.1_vae_bf16.safetensors	vae	644.2 MiB (675509688 bytes)	bb21f7473051e1ac368515dd3f2e15cd44d7a11748ee8823e1ddca3e4876b7c9	Comfy-Org/Qwen-Image-2.1
Qwen3VL-8B-Instruct-Q4_K_M.gguf	text-encoder	4.68 GiB (5027784800 bytes)	67d1659bfe71b89d50b45a4ad1a9e5b997e5bb16ce5da66a6a6167abd569e9e2	Qwen/Qwen3-VL-8B-Instruct-GGUF
mmproj-Qwen3VL-8B-Instruct-F16.gguf	mmproj	1.08 GiB (1159029824 bytes)	ca524100ebf825c9a870db1c580d03879e0da0ab2541697e2458e64891cf9d38	Qwen/Qwen3-VL-8B-Instruct-GGUF

Total download: 10.30 GiB (11061889336 bytes).

q8

File	Role	Size	SHA-256	Origin
qwen-image-2.1-Q8_0.gguf	model	7.12 GiB (7640860384 bytes)	c0ed4b2ffd56cbe9c3df1e4a4098045256484ebe94ba5a7e4338d35ec046baa5	unsloth/Qwen-Image-2.1-GGUF
vae/qwen_image_2.1_vae_bf16.safetensors	vae	644.2 MiB (675509688 bytes)	bb21f7473051e1ac368515dd3f2e15cd44d7a11748ee8823e1ddca3e4876b7c9	Comfy-Org/Qwen-Image-2.1
Qwen3VL-8B-Instruct-Q4_K_M.gguf	text-encoder	4.68 GiB (5027784800 bytes)	67d1659bfe71b89d50b45a4ad1a9e5b997e5bb16ce5da66a6a6167abd569e9e2	Qwen/Qwen3-VL-8B-Instruct-GGUF
mmproj-Qwen3VL-8B-Instruct-F16.gguf	mmproj	1.08 GiB (1159029824 bytes)	ca524100ebf825c9a870db1c580d03879e0da0ab2541697e2458e64891cf9d38	Qwen/Qwen3-VL-8B-Instruct-GGUF

Total download: 13.51 GiB (14503184696 bytes).

Notes

• The Qwen Research License is not Apache-2.0; review its restrictions before use, particularly for commercial deployments.
• The default official INT8 ConvRot pipeline pairs the Qwen3-VL-8B-Instruct Q4_K_M encoder with the model's BF16 VAE and F16 vision projector. The projector is required for edits; older Qwen Image VAEs cannot replace the 2.1 VAE.
• Upstream supports ten reference images, but this API caps editing at eight. See the Image API for multipart uploads and remote reference URLs.
• Use dimensions divisible by 32. Request transparent RGBA output in the prompt and select PNG to preserve alpha.
• CPU and CUDA builds use the same official server. CPU inference can be slow; total working memory exceeds downloaded weight sizes.
• Registration and mocked protocol tests are not an end-to-end inference benchmark. Validate generation and editing on target hardware.

Quality benchmarks

Arena Text-to-Image: #19 (1222.93)

qwen-image-2.1

Upstream model; not this local quantization

Published 2026-10-07; 11441 votes

Checked 2026-10-13

Arena Image Edit: #18 (1368.69)

qwen-image-2.1

Upstream model; not this local quantization

Published 2026-10-06; 21941 votes

Checked 2026-10-13

Benchmarks

No submitted benchmark reports yet.