Qwen team releases Qwen Image 2.1 Turbo, an 8-step checkpoint under a non-commercial licence

The Qwen team released Qwen-Image-2.1-Turbo on October 9, 2026, an accelerated checkpoint of its Qwen Image 2.1 model that generates and edits images in 8 denoising steps, where the original model’s documented default is 40. The team says Turbo uses the same 7B visual generation architecture as the original, which the team released on September 20. The weights are on Hugging Face under the Qwen Research License Agreement, not an open source licence, and that agreement grants use for non-commercial purposes only. Hugging Face lists 32.44 GB of weight files for Turbo. We have not run it, and the Qwen team has published no speed or quality figure for Turbo against the original.

Takeaway points

• The Qwen team released Qwen-Image-2.1-Turbo on October 9, 2026, a 7B image generation and editing checkpoint that runs in 8 denoising steps against a default of 40 for the original model.
• The Qwen Research License Agreement grants rights for non-commercial purposes only and says commercial use needs a separate licence from the company.
• The Qwen team has published no timing or quality comparison between Turbo and the original model, and we have not run Turbo on our test machine.

What is Qwen Image 2.1 Turbo, and what changed from the original?

Qwen Image 2.1 Turbo is a second set of weights for the same model design, made to finish an image in fewer steps. A diffusion model starts from noise and removes it over a series of steps; each step is one pass through the network. The model card describes Turbo as “an accelerated checkpoint” for text-to-image generation and image editing, and the code repository’s default parameters table lists 40 inference steps for the original checkpoint. Hugging Face records the original as the base model of the Turbo repository.

The visual generation part, which the team calls the transformer, is the same size in both repositories. Hugging Face counts 7,115,124,736 parameters in 16-bit (BF16) format for each, and the two transformer files are byte for byte the same size, 9,968,332,504 and 4,261,951,904 bytes. Their checksums differ, so the weights inside them are new. The code repository describes this part as a single-stream model with 32 layers and 7B parameters.

What makes the 8 steps is a fixed schedule saved with the checkpoint. Turbo’s pipeline file carries a list of eight noise levels, from 1.0 down to 0.414568, which the original’s file does not have. Its scheduler file also switches off dynamic shifting, which the original’s scheduler file switches on. The model card says generation uses a guidance setting of 1 by default, and that prefix KV caching reuses the text and reference image context across the steps.

The image decoder, the VAE, holds the same 337,740,404 parameters in both repositories, by our count of the file headers. The original stores them in 32-bit numbers and Turbo in 16-bit, which is why Turbo’s decoder file is half the size and its total download is 0.68 GB smaller.

Can Qwen Image 2.1 Turbo be used commercially?

No, not under the licence that comes with the weights. The Qwen Research License Agreement, dated September 20, 2026, grants rights to use, copy, modify and share the materials for non-commercial purposes only, and it defines non-commercial as “for research or evaluation purposes only”. Section 2 says the materials may not be used for any commercial purpose without a separate commercial licence from the company, and names an address to request one. The licensor is Hangzhou Tongyi Laboratory Technology Co., Ltd.

The same licence file sits in both repositories, and the two copies are identical. Hugging Face lists the licence as “other” with the name qwen-research, and neither repository is gated, so the files download without an access request.

The agreement sets conditions beyond the commercial limit. Anyone who shares the weights or a derivative must pass on a copy of the agreement, mark changed files, and keep a notice text crediting the licensor. Anyone who uses the materials or their outputs to build or improve an AI model that is distributed must display “Built with Qwen” or “Improved using Qwen” in its documentation. “Qwen” may not be the main name of a derivative product. The licence ends for anyone who brings an intellectual property claim against the company over the materials or their output. It is governed by the laws of China, with disputes going to the courts in Hangzhou.

On the day it released the weights, the team also said hosted Qwen-Image-2.1 Pro and Turbo programming interfaces were “now officially live” on Alibaba Cloud Model Studio. The code repository offers the two routes side by side: the hosted interfaces, or the Turbo checkpoint run with the Diffusers library. Its pages do not set out the terms of the hosted service, and we did not look at them.

How big is the Qwen Image 2.1 Turbo download, and will it fit a 48 GB Mac?

The weight files come to 32.44 GB, which is 15.56 GB less than the 48 GB of unified memory on our test machine, an Apple M4 Pro. With the example images, tokenizer and other files, the whole repository is 32.49 GB. Fitting the files is not the same as running them, and we have not tried. A Mac also needs memory for the operating system and for the work of each step, and what else is loaded matters too, as our test of two models at once on a 48 GB Mac showed.

The 7B figure covers only the transformer. The download also holds a text encoder, which the code repository names as Qwen3-VL 8B, a vision and language model that reads both the prompt and any reference images. It is one 17.53 GB file, and by our count of its header it has 8,767,123,696 parameters. Add the transformer and the image decoder and the three parts hold 16,219,988,836 parameters, by our arithmetic.

The official instructions are written for Nvidia hardware. The model card begins with “Install a CUDA-compatible PyTorch build”, and both of its code examples move the pipeline to a CUDA device. The code repository added a section for Apple silicon on October 14. It points to a native macOS app from the LightX2V project, a third party, and says it can run Turbo on a Mac for text-to-image and image-to-image work. The Qwen team gives no speed or memory figure for that route. Apple’s MLX format, the one covered in our piece on what MLX is in LM Studio, is not among the formats in the official repository, which holds safetensors files only.

How do the 8 steps in Qwen Image 2.1 Turbo work, and can they be changed?

The 8 steps come from the saved schedule, and the usual step setting does not change them. The model card says “Setting num_inference_steps alone does not override it.” A user can pass a different list of noise levels when calling the pipeline, but the card adds that “other schedules have not been evaluated for this checkpoint.”

The saved schedule needs a version of Diffusers that the team asks users to install from source. The model card names the change that added support, a Diffusers pull request titled “Configure fixed sampling sigmas on QwenImage21Pipeline”, which was merged on October 5, four days before the release. The card also asks for version 5.17.0 or later of the Transformers library.

Turbo uses the same seven resolution presets as the original, from 2048 by 2048 pixels square to 2752 by 1536 for a 16:9 frame. The card’s text-to-image example renders a 1696 by 2528 portrait and its editing example a 2048 by 2048 square. The original model, according to its card, accepts up to 10 reference images for editing and makes transparent images with an alpha channel; the Turbo card’s showcase includes transparent image generation and a four-image composition. For larger models that do not fit, quantized builds such as the 58.41 GB Qwen3.8 Flash Next GGUF build from DASLab are one route others have taken, but no such build of Turbo is in the official repository.

What has the Qwen team not published about Qwen Image 2.1 Turbo?

The team has not published how much faster Turbo is, or how its images compare with the original’s. Neither the Turbo model card nor the code repository gives a timing, a memory figure or a quality score for Turbo. The card’s showcase has one example image in each of eight categories, chosen by the team. The step count is the only figure on the record. Our reading is that 8 steps means fewer passes through the transformer, while the text encoder and the image decoder still do their work once per image, so the time saved per image cannot be read from the step count alone.

The team has also not said how the faster checkpoint was trained. The documents call it accelerated and give its schedule, and they do not name a method. The Qwen-Image-2.1 blog post that the model card links to did not load for us: the page returned a script shell to a plain request and to our page reader, so we have not used it.

Two details in the team’s own documents do not line up. The code repository’s architecture section says the model’s scheduler uses “dynamic shifting”, which is true of the original’s scheduler file and not of Turbo’s, which turns it off; the section was written for the original release. The news list in the same repository dates Intel’s support through its OpenVINO toolkit to September 20, while the change that added that line was made on October 13.

On October 15 Hugging Face showed 15,315 downloads and 824 likes for the Turbo repository, against 159,495 downloads and 3,316 likes for the original.

We have not run Qwen-Image-2.1-Turbo or the original model on our Apple M4 Pro with 48 GB, so this piece holds no speed, memory or image quality figure of our own. Every figure above comes from the Qwen team’s model cards, code repository, licence file and the Hugging Face data for the two repositories, saved on October 15, 2026.