Qwen/Qwen-Image-2.1-Turbo 7B

• License: Qwen Research License Agreement
• Modalities: Input: text, image, multiple images; Output: image
• Original model: Qwen/Qwen-Image-2.1-Turbo

Description

Qwen-Image-2.1-Turbo accelerates the Qwen-Image-2.1 7B visual generation architecture for eight-step text-to-image generation and reference-image editing, including transparent images. The separate text encoder is not included in the 7B parameter count.

Run

self serve qwen-image-2.1-turbo:7b

See the Image API for generation and editing requests.

Example Request

curl http://localhost:8080/v1/images/generations \
        -H 'Content-Type: application/json' \
        -d '{"prompt":"A red bicycle leaning against a white wall","size":"1024x1024","n":1,"output_format":"png"}' \
        -o response.json
jq -r '.data[0].b64_json' response.json | base64 --decode > output.png

Engine Parameters

Parameter	Default	What it controls
steps	8	Number of denoising steps.
sigmas	1.0,0.978453,0.95418,0.926626,0.89508,0.845148,0.704534,0.414568,0.0	Published Turbo schedule with terminal zero.
cfg_scale	1.0	Guidance scale.
sampling_method	euler	Sampling method.
flash_attention	true	Enable flash attention.
offload_to_cpu	true	Offload inactive model stages to CPU.

Override startup settings with --set <parameter>=<value>. Per-request sampling overrides are documented in the Image API.

Notes

Both quantizations require the Qwen Image 2.1 BF16 VAE, Qwen3-VL-8B-Instruct text encoder and F16 vision projector listed in the generated registry details. The API accepts at most eight reference images. Use dimensions divisible by 32 and PNG output to preserve transparency.

The checkpoint-specific eight-step sigma schedule is passed to sd-server with --sigmas.

Local Validation

On 2026-10-11, both registered Q4 and Q8 variants were downloaded through self pull with the source registry and verified against its artifact pins. Both started through self serve --device cuda on an NVIDIA RTX 5090 using stable-diffusion.cpp master-929-3f8527a and the registry sampling defaults.

• Q4 and Q8: health returned HTTP 200; generation returned one valid 512x512 PNG for "A simple bright red apple on a white background" (seed 7). Visual inspection confirmed the requested object, color and background.
• Q4: a single-reference edit changed the apple to green while preserving its shape, stem and composition, returning a valid 512x512 PNG.
• Q4: an empty generation prompt returned HTTP 400 invalid_request.

These are basic inference smoke tests. Q8 editing, multiple-reference edits, transparent output, CPU execution and equivalence to the official Diffusers pipeline remain unverified.

Registry details

• Registry id: qwen-image-2.1-turbo:7b
• Type: image
• Default quantization: q4
• Capabilities: text, image, multi-image -> image, image-edit

Benchmarks

No submitted benchmark reports yet.

Available quants

Quant	Adapter	Repository	File	Size	SHA-256
q4 (default)	sd-server	unsloth/Qwen-Image-2.1-Turbo-GGUF	qwen-image-2.1-turbo-Q4_K_M.gguf	3.9 GiB	8940e8b522af886e0c0f5afddf720e82dd223cb9a43580b5f61c4d740452ac20
			vae/qwen_image_2.1_vae_bf16.safetensors	644 MiB	bb21f7473051e1ac368515dd3f2e15cd44d7a11748ee8823e1ddca3e4876b7c9
			Qwen3VL-8B-Instruct-Q4_K_M.gguf	4.7 GiB	67d1659bfe71b89d50b45a4ad1a9e5b997e5bb16ce5da66a6a6167abd569e9e2
			mmproj-Qwen3VL-8B-Instruct-F16.gguf	1.1 GiB	ca524100ebf825c9a870db1c580d03879e0da0ab2541697e2458e64891cf9d38
q8	sd-server	unsloth/Qwen-Image-2.1-Turbo-GGUF	qwen-image-2.1-turbo-Q8_0.gguf	7.1 GiB	399d2ddc3636f36456e2f5bfa57187da4e38721bc046430317f42a9eab5b1712
			vae/qwen_image_2.1_vae_bf16.safetensors	644 MiB	bb21f7473051e1ac368515dd3f2e15cd44d7a11748ee8823e1ddca3e4876b7c9
			Qwen3VL-8B-Instruct-Q4_K_M.gguf	4.7 GiB	67d1659bfe71b89d50b45a4ad1a9e5b997e5bb16ce5da66a6a6167abd569e9e2
			mmproj-Qwen3VL-8B-Instruct-F16.gguf	1.1 GiB	ca524100ebf825c9a870db1c580d03879e0da0ab2541697e2458e64891cf9d38