Qwen-Image-2.1-Turbo 8-Step Fast Image Generation Model: 48GB+ VRAM


- 1. At a Glance
- 2. Overview
- 3. Specifications
- 4. Performance and Quality
- 5. Strengths and Use Cases
- 6. How It Differs from Similar Models
- 7. Hardware Requirements
- 8. Can You Run It Locally?
- 9. Our Own Measurements
- 10. Recent Models in the Same Size Class
- 11. Distributed Files
- 12. How to Get It
- 13. Quantized and Converted Variants
- 14. Related Articles
- 15. What to Read Next
- 16. Sources
At a Glance
| Item | Value |
|---|---|
| Repository | Qwen/Qwen-Image-2.1-Turbo |
| Family guide | Qwen-Image-2.1 guide (5 articles) |
| Publisher guide | Alibaba (Qwen): models and licenses |
| Published | 2026-10-09 |
| License | qwen-research |
| Formats | safetensors |
| Source type | Primary source (the publisher itself) |
Values determined by this site’s code when the information was collected. Dates are JST.
Overview
The Qwen team has released “Qwen-Image-2.1-Turbo", an open-weight model supporting text-to-image generation and image editing.
This model is a speed-optimized checkpoint based on the previously released “Qwen-Image-2.1". By providing text prompts or input images, it can generate high-quality images in a mere 8 steps of denoising processing.
Specifications
The main specifications listed in the release materials are as follows:
- Parameter Count & Architecture: Employs 7B parameters for the visual generation component. Like the base Qwen-Image-2.1 model, it inherits a 32-layer Single-Stream DiT (Diffusion Transformer) architecture and mixed-granularity attention.
- Inference Steps: 8 steps (the recommended sampling schedule is saved within the checkpoint in advance and loaded automatically).
- Guidance Scale: Uses CFG (Classifier-Free Guidance) = 1 by default.
- Context Processing: Supports prefix KV caching, efficiently reusing text and reference image context between denoising steps.
- Output Resolution: Supports the 2048 resolution level. Specific resolution presets by aspect ratio are as follows:
- 1:1: 2048 × 2048 – 4:3: 2400 × 1792 – 3:4: 1792 × 2400 – 3:2: 2528 × 1696 – 2:3: 1696 × 2528 – 16:9: 2752 × 1536 – 9:16: 1536 × 2752
- Input & Editing Features: In addition to generating new images from text, it supports instruction-based image editing using specified input images.
- License: Governed by the qwen-research license.
Performance and Quality
While the release materials do not contain an objective numerical benchmark comparison table, the publishers emphasize the model’s ability to maintain high expressiveness with a low count of just 8 steps.
Whereas the base Qwen-Image-2.1 model standardly required 40 steps, this Turbo model achieves a drastic reduction in inference time by using a specially tuned 8-step sampling schedule. Furthermore, because it is optimized for generation at CFG=1, it does not require additional passes associated with negative prompt calculations.
According to the showcase categories presented by the publishers, it is reported that a wide range of expression and editing tasks can be covered in 8 steps, including:
- Portrait photography (natural human depiction and lighting)
- Depiction of complex human poses and dynamic movements
- Generation of transparent images (RGBA) including alpha channels
- Typography and poster design (accurate text rendering and layout)
- UI design and the rendering of diagrams and information layouts
- Style and texture conversion from a single reference image (such as converting sketches to photorealistic photos)
- Compositional synthesis combining multiple reference images, as well as multi-view interior composition
A major strength of this model is that it delivers practical outputs at a low computational cost of just 8 steps without sacrificing the delicate detail rendering, accurate text rendering, and flexible image editing capabilities possessed by the base Qwen-Image-2.1 model.
Strengths and Use Cases
Qwen-Image-2.1-Turbo is a checkpoint that achieves fast inference in just 8 steps while maintaining the general-purpose generation and editing capabilities of the base Qwen-Image-2.1 model. Therefore, it is suited for production environments that require rapid iteration of generations and for interactive image generation workflows.
The specific domains of strength and expected use cases for this model are as follows:
- High-Resolution Text-to-Image Generation
It supports diverse aspect ratios including 2048×2048 (1:1), 16:9 (2752×1536), and 9:16 (1536×2752), and can directly output high-definition images at the 2048 level. It is suited for expressing portrait photography, fine poses and movements of people, and complex scene descriptions with sharp textures. - Precise Typography and UI/Diagram Layouts
It boasts strengths in rendering text within images (typography), poster production, UI mockups, and generating information layouts such as flowcharts and recipe diagrams. It can arrange detailed text placement and diagram elements specified in prompts without breaking down. - Direct Generation of Transparent Images (RGBA)
It can directly output background-transparent images containing alpha channels. It is suited for use cases where illustration assets, stamp images, or product cutouts need to be generated without post-processing. - Diverse Image Editing and Conversion
It supports single-image style conversion that turns hand-drawn sketches into photorealistic photos, compositional synthesis combining multiple reference images, and indoor spatial layout configuration. Editing can be performed while efficiently reusing reference image context through prefix KV caching.
How It Differs from Similar Models
As derivative models based on Qwen-Image-2.1, there are multiple models that have been covered on our site in the past. There are differences between those and this model as follows:
First, as another model aiming for acceleration, there is the distillation model published by Viggle 「Qwen-Image-2.1-viggle-turbo」4ステップの高速画像生成・編集モデル:必要VRAM 48GB〜. While the Viggle version pursues generation in just 4 steps (4 passes) using Distribution Matching Distillation (DMD), the current Qwen-Image-2.1-Turbo is an 8-step accelerated checkpoint provided by the official Qwen team. In contrast to the Viggle version which targets extreme step reduction, the official Turbo adopts 8 steps and a dedicated sampling schedule, designed to strike a balance between generation speed and meticulous detail rendering.
Additionally, as an adapter specialized for specific tasks, there is 「QI_2.1_AnyAngle」カメラアングル操作LoRA:必要VRAM 48GB〜・配布ファイル一覧 which enables free control of camera angles. While QI_2.1_AnyAngle is a LoRA extension that transforms viewpoints in combination with rough references like 3D models, Qwen-Image-2.1-Turbo is not specialized for specific angle operations, but is positioned as an alternative base model checkpoint that condenses the wide range of functions possessed by the base model—such as text generation, image editing, and transparent image generation—entirely into 8 steps.
Hardware Requirements
Estimated requirements (calculated by Local Model Watch) — 7.1B parameters
| Your VRAM | Quantization | File size | Est. memory needed |
|---|---|---|---|
| 48GB (RTX 6000 Ada / A6000, etc.) | BF16 | 30.2GB | 36.3GB |
Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.
Can You Run It Locally?
The publisher distributes this model as safetensors.
License — qwen-research (Commercial use prohibited): Alibaba’s Qwen Research License. Limited to research and evaluation; commercial use requires a separate license (Sections 1.i and 2.b).
Compiled by this site’s code from the published formats and the license field. License summaries are not legal advice — check the publisher’s original terms before relying on them.
Our Own Measurements
Measurements We Did Not Take
This model’s license (qwen-research) does not permit commercial use. Because this site carries advertising, we did not run the model (no CPU run, answers, quantization comparison, conversion or generation).
Recent Models in the Same Size Class
Models with 4–15B parameters that Local Model Watch covered recently, listed by code from our article log for comparison. VRAM tiers are this site’s estimates; licenses are as stated on the model cards.
→ Scroll horizontally to see all columns
| Model | Parameters | Smallest VRAM tier | License | Our article |
|---|---|---|---|---|
| lodestones/Kromoss | 5.1B | 32GB | apache-2.0 | Kromoss 1024px Image Generation Model: 32GB+ VRAM, File List (2026-11-17) |
| briaai/Fibo-2.0-dev | 8.1B | 32GB | bria-fibo | Fibo-2.0-dev Image Generation Model: 32GB+ VRAM, File List (2026-11-17) |
| unsloth/MOSS-Image-SFT-GGUF | 5.1B | 32GB | apache-2.0 | MOSS-Image-SFT-GGUF 5.1B Text-to-Image Generation Model: 32GB+ VRAM (2026-11-17) |
Distributed Files
Weight files published in Qwen/Qwen-Image-2.1-Turbo, listed by this site from the Hugging Face API. Sizes are the actual file sizes.
| File | Size |
|---|---|
text_encoder/model.safetensors |
17.53GB |
transformer/diffusion_pytorch_model-*-of-00002.safetensors |
14.23GB (2 split files) |
vae/diffusion_pytorch_model.safetensors |
676MB |
How to Get It
Qwen-Image-2.1-Turbo is available on Hugging Face in safetensors format. Since it is not a gated model, it can be downloaded and used directly without individual agreement to terms of use or the like.
Using this model requires the latest development version of the Diffusers library. Because a feature to load the sampling sigmas set in the checkpoint on the pipeline side (implemented in Diffusers PR #14950) is required, you need to install the latest source code directly from the Diffusers repository.
The command to install the required packages is as follows:
pip install git+https://github.com/huggingface/diffusers.git
pip install "transformers>=5.17.0" accelerate pillow
Construct the pipeline from Python code using QwenImage21Pipeline from Diffusers.
import torch
from diffusers import QwenImage21Pipeline
pipe = QwenImage21Pipeline.from_pretrained(
"Qwen/Qwen-Image-2.1-Turbo",
dtype=torch.bfloat16,
).to("cuda")
image = pipe(
prompt="Your text prompt here",
width=2048,
height=2048,
use_kv_cache=True,
generator=torch.Generator("cpu").manual_seed(20261008),
).images[0]
image.save("output.png")
The recommended 8-step sampling schedule is saved in advance within the checkpoint and is applied automatically upon loading. Therefore, it is not necessary to explicitly set the scheduler in the pipeline arguments. Note that specifying num_inference_steps in the arguments will not overwrite the saved schedule, and when experimenting with changing the schedule, sigmas must be explicitly specified at call time.
Quantized and Converted Variants
→ Scroll horizontally to see all columns
| Added | Publisher | Format | Repository | Smallest VRAM tier (build, est. memory) |
|---|---|---|---|---|
| 2026-10-11 | unsloth | GGUF | unsloth/Qwen-Image-2.1-Turbo-GGUF | BF16 36.3GB (fits in 48GB VRAM) |
| 2026-10-11 | unsloth | FP8 | unsloth/Qwen-Image-2.1-Turbo-FP8 | BF16 36.3GB (fits in 48GB VRAM) |
File sizes of each build:
- Available builds in unsloth/Qwen-Image-2.1-Turbo-GGUF: BF16 30.2GB
- Available builds in unsloth/Qwen-Image-2.1-Turbo-FP8: BF16 30.2GB
In addition, 48 converted build(s) from other uploaders exist on Hugging Face; this site lists only builds from the model’s publisher or established quantization maintainers.
This section is appended automatically by Local Model Watch when a converted build of this model appears after publication. Memory figures are estimated from the size of the distributed files. What the quantization names mean: glossary.
Related Articles
- Qwen-Image-2.1-viggle-turbo Image Generation Model: 48GB+ VRAM
- PrunaAI Releases Object Removal LoRA for Qwen-Image-2.1
- QI_2.1_AnyAngle Camera Angle Control LoRA: 48GB+ VRAM, File List
- Qwen-Image-2.1 Released: Open-Weight Image Gen & Editing
What to Read Next
- Find models by VRAM (This model needs at least 48GB) → VRAM quick reference
- Explore the same model family → Qwen-Image-2.1 family overview (5 articles, 2 converted builds)
- Formats this model is available in → GGUF format guide and models / FP8 format guide and models
- Learn about the publisher → Alibaba (Qwen): models, licenses and articles
- Other models for the same task → Other image generation models

