Last updated: October 13, 2026

Qwen Image 2.1 Turbo: what it is, what it costs, how to use it

On October 9, 2026 Alibaba’s Qwen team released Qwen-Image-2.1-Turbo, an accelerated checkpoint of the open-weight Qwen-Image-2.1 that came out on September 20. It generates and edits images in 8 denoising steps instead of 40, on the same 7B architecture. This page covers what changed, where to run it, what it costs and the licence catch.

Qwen-Image-2.1-Turbo at a glance

Source: Hugging Face model card

Source: Qwen-Image-2.1 on GitHub

Turbo vs the base model vs Pro

Turbo is not a smaller model. It is the same 7B network trained to finish in fewer steps, so each image needs about a fifth of the denoising passes. Pro is the hosted top tier on Alibaba Cloud.

The base model’s 60.28 was measured on the 40-step checkpoint. Do not assume Turbo matches it until Qwen or an independent tester publishes a Turbo score.

Source: Hugging Face model card

Source: Model Studio: qwen-image-2.1-turbo

Source: Model Studio: qwen-image-2.1-pro

Source: DataNorth: Qwen-Image-2.1 benchmark chart

Qwen Image 2.1 Turbo price

Alibaba prices are list prices before promotions. Runware and Venice are third-party hosts with their own terms. See the API guide for code and the details of each.

Qwen Image 2.1 Turbo API guide

Source: Model Studio: qwen-image-2.1-turbo

Source: Runware model page

Source: Venice model page

How to use Qwen Image 2.1 Turbo

• Try it in a browser. The Qwen-Image-2.1 demo Space on Hugging Face runs the model family without setup. Venice also offers it with a free daily allowance for new accounts.
• ComfyUI. Load the official Qwen-Image 2.1 workflow, swap in the Turbo diffusion model (or the base model plus the Turbo LoRA), and set 8 steps and CFG 1.
• Python (Diffusers). QwenImage21Pipeline loads the checkpoint directly. You need Diffusers from source and transformers 5.17 or newer.
• Hosted API. Call qwen-image-2.1-turbo on Alibaba Cloud Model Studio, or use a third-party host such as Runware.

ComfyUI and local install guide · API guide

Source: Qwen-Image-2.1 demo Space

Source: Venice model page

Source: ComfyUI Wiki release note

Source: Hugging Face model card

Quick start in Python

Install the libraries, then generate an image. The pipeline reads the 8-step schedule from the checkpoint, so you do not pass a step count.

pip install "torch>=2.4.0"
pip install git+https://github.com/huggingface/diffusers.git
pip install "transformers>=5.17.0" accelerate pillow

import torch
from diffusers import QwenImage21Pipeline

pipe = QwenImage21Pipeline.from_pretrained(
    "Qwen/Qwen-Image-2.1-Turbo", dtype=torch.bfloat16
).to("cuda")

# The 8-step schedule and CFG 1 are stored in the checkpoint: no steps argument needed.
image = pipe(
    prompt="A minimalist coffee shop poster, the words 'OPEN LATE' in bold serif type, warm evening light",
    width=1696,
    height=2528,
    use_kv_cache=True,
    generator=torch.Generator("cpu").manual_seed(42),
).images[0]
image.save("t2i.png")

Setting num_inference_steps alone does not change the schedule; only an explicit sigmas argument overrides it, and Qwen says other schedules are untested.

Source: Hugging Face model card

Image editing and multi-reference

The same pipeline edits images: pass image= with an instruction prompt. Pass a list to combine up to 10 reference images.

import torch
from diffusers import QwenImage21Pipeline
from diffusers.utils import load_image

pipe = QwenImage21Pipeline.from_pretrained(
    "Qwen/Qwen-Image-2.1-Turbo", dtype=torch.bfloat16
).to("cuda")

sketch = load_image("sketch.png").convert("RGBA")
image = pipe(
    prompt="Turn this sketch into a photorealistic yacht on the open sea at golden hour",
    image=sketch,  # pass a list for multi-reference editing (up to 10 images)
    width=2048,
    height=2048,
    use_kv_cache=True,
    generator=torch.Generator("cpu").manual_seed(20261009),
).images[0]
image.save("edit.png")

Source: Hugging Face model card

Source: Qwen-Image-2.1 on GitHub

Recommended resolutions

For transparent PNGs, Qwen recommends stating in the prompt that the image has transparency and a transparent background.

Source: Hugging Face model card

Source: Qwen-Image-2.1 on GitHub

Benchmarks: what Qwen has and has not published

Qwen published only one chart for the 2.1 family, on its own Qwen-Image-Bench, which is not public. All scores are for the 40-step base model.

• No Turbo score, latency or throughput figure has been published by Qwen.
• Qwen gives no per-category breakdown or methodology for the chart.
• Runware lists about 23 seconds per request on its own hosted setup; that includes queueing and is not a GPU benchmark.

Source: DataNorth: Qwen-Image-2.1 benchmark chart

Source: Runware model page

Can you use it commercially?

The open weights are under the Qwen Research License Agreement, the same as the base model. Research and personal experiments are fine; running it for a business on your own servers needs a separate agreement with Alibaba. Images made through Alibaba Cloud Model Studio or a paid third-party host fall under that service’s terms instead, so read them before you ship.

Source: Hugging Face model card

Source: Qwen-Image-2.1 on GitHub

Independent guide, not affiliated with Alibaba or the Qwen team. Prices, licences and file names change; check the linked official pages before you build on them.