Alibaba’s Qwen crew has launched Qwen-Picture-2.1-Turbo, an accelerated checkpoint of its open-weight Qwen-Picture-2.1 mannequin. It generates and edits images in 8 denoising steps as a substitute of the bottom mannequin’s 40-step default. For builders, meaning 5x fewer denoising steps on the identical 7B structure, plus a hosted API choice.
TL;DR
- Measurement: 7B parameters within the visible generator, paired with a Qwen3-VL 8B text encoder.
- Runs on: CUDA GPUs in BF16 by way of Diffusers. Qwen publishes no Turbo VRAM minimal. Unsloth estimates the bottom mannequin runs on 11 GB VRAM with GGUF and 24 GB with INT8/FP8.
- Efficiency: Identical 2K output and modifying function set as Qwen-Picture-2.1, at 8 steps as a substitute of 40.
- Finest: Base Qwen-Picture-2.1 scores 60.28 on Qwen-Picture-Bench, the highest open-weight rating Qwen studies.
- Backside line (finest): 8-step 2K technology and modifying in a single open checkpoint.
- Backside line (worst): Analysis license solely, so industrial self-hosting wants separate permission.
What’s Qwen-Picture-2.1-Turbo?
Qwen-Picture-2.1-Turbo is an 8-step picture technology and modifying checkpoint from Alibaba’s Qwen crew. It’s constructed on Qwen-Image-2.1. Turbo retains the identical 7B visible technology structure. It hundreds instantly with QwenImage21Pipeline in Diffusers.
The checkpoint ships with its beneficial 8-step sampling schedule saved inside. Technology makes use of CFG=1 by default. Prefix KV caching reuses the textual content and reference-image context throughout denoising steps.
How does Qwen-Picture-2.1-Turbo work?
The underlying architecture is a single-stream DiT with 32 layers and 7B parameters. It makes use of block-causal consideration. Textual content tokens get a token-level causal masks, whereas pictures use a chunk-level bidirectional masks.
- Textual content encoder: Qwen3-VL 8B encodes each directions and situation pictures.
- VAE: a 64-channel RGBA autoencoder with 16x spatial compression, which permits native transparency.
- Scheduler: Circulate Matching with Euler discrete scheduling and dynamic shifting.
The eye design is what makes prefix caching work. Enter pictures and textual content are computed as soon as at step one. Each later step reuses that cache. With solely 8 steps, the cached prefix covers a lot of the conditioning price.
What can it generate and edit?
The Turbo model card showcase covers 8 classes. These embrace portraits, human poses, clear pictures, typography and posters, and UI layouts. Modifying examples embrace single-image transformation, multi-reference composition and 4-image inside composition. The bottom mannequin helps as much as 10 reference images and native edits by way of circles, painted annotations or masks.
How do you run Qwen-Picture-2.1-Turbo?
Setup requires Diffusers from source, together with transformers>=5.17.0. The checkpoint wants Diffusers PR #14950, which provides pipeline-configured sampling sigmas.
import torch
from diffusers import QwenImage21Pipeline
pipe = QwenImage21Pipeline.from_pretrained(
"Qwen/Qwen-Image-2.1-Turbo", dtype=torch.bfloat16
).to("cuda")
picture = pipe(
immediate="A ceramic teapot on a wooden table, soft window light",
width=2048, top=2048, use_kv_cache=True,
).pictures[0]
For modifying, go picture=input_image with an instruction immediate. Supported presets run from 2048×2048 sq. to 2752×1536 at 16:9.
One factor to notice right here. Setting num_inference_steps alone does not override the saved schedule. Solely an express sigmas argument does, and Qwen notes different schedules are untested.
What does the API price?
Alibaba Cloud Mannequin Studio now hosts each Turbo and Professional. qwen-image-2.1-turbo prices CNY 0.1 per picture in most areas, with a 120 RPM restrict. qwen-image-2.1-pro prices CNY 0.25 per picture, with 20 RPM. Turbo is 2.5x cheaper per picture and permits 6x the request price.
Qwen-Picture-2.1-Turbo vs competing quick picture fashions
| Characteristic | Qwen-Image-2.1-Turbo | Qwen-Image-2.1 | Z-Image-Turbo | FLUX.2 klein 9B |
|---|---|---|---|---|
| Org | Alibaba Qwen | Alibaba Qwen | Alibaba Tongyi-MAI | Black Forest Labs |
| Generator measurement | 7B | 7B | 6B | 9B |
| Textual content encoder | Qwen3-VL 8B | Qwen3-VL 8B | Not disclosed | Qwen3 8B |
| Default steps | 8 | 40 | 8 NFEs | 4 |
| Modifying | Sure, multi-reference | Sure, as much as 10 references | No (separate Edit mannequin) | Sure, multi-reference |
| Clear RGBA output | Sure | Sure | Not disclosed | Not disclosed |
| Native decision | 2K (2048×2048) | 2K (2048×2048) | 1024×1024 in examples | 1024×1024 in examples |
| {Hardware} steerage | Not disclosed | 11 GB GGUF / 24 GB FP8 (Unsloth) | Matches 16 GB VRAM | ~29 GB VRAM, RTX 4090+ |
| License | Qwen Analysis License | Qwen Analysis License | Apache 2.0 | FLUX Non-Industrial |
| Qwen-Picture-Bench | Not disclosed | 60.28 (vendor) | Not disclosed | Not disclosed |
| Hosted API | CNY 0.1/image | Professional: CNY 0.25/image | Not disclosed | BFL API |
Key Takeaways
- Qwen-Picture-2.1-Turbo cuts denoising from 40 steps to eight on the identical 7B structure.
- One checkpoint handles 2K text-to-image, multi-reference modifying and clear RGBA output.
- The API prices CNY 0.1 per picture, 2.5x cheaper than Professional.
- Weights are research-licensed, so industrial self-hosting wants separate permission.
- No Turbo-specific benchmark exists but; base Qwen-Picture-2.1 scores 60.28 on Qwen-Picture-Bench.
Take a look at the ModelScope page, Hugging Face page, Pro API and Turbo API. All credit score goes to the researcher of this undertaking. Additionally, be happy to observe us on Twitter and don’t neglect to affix our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.
The publish Alibaba Qwen Releases Qwen-Image-2.1-Turbo, an 8-Step 7B Image Model appeared first on MarkTechPost.

