Qwen-Image 2.1 Runs Locally: One Open Model That Draws and Edits

Alibaba's Qwen team published Qwen-Image 2.1 on 20 September 2026, and the interesting part is not the numbers. It is that there is only one model now. Text to image and image editing used to be two downloads with two sets of nodes. In 2.1 they are the same weights, and the difference is whether you hand it a reference image or not.

It is also a smaller download than its predecessor, which is the opposite of what people expect from a new release.

The short version

Qwen-Image 2.1 is worth having on a 24 GB card if you are allowed to use it. The licence is the catch, and it is a real one: this is a research licence, not Apache 2.0, so commercial work needs either a separate agreement with Qwen or one of the older Apache models instead. Read the licence section below before you build anything on it.

A short timeline

• August 2025, Qwen-Image (2508). The first one. 20 billion parameters in the image tower, Qwen2.5-VL as the text encoder, and a separate Qwen-Image-Edit line for editing.
• February 2026, Qwen-Image 2.0. Native 2K output.
• 20 September 2026, Qwen-Image 2.1. 7 billion parameters in the image tower, a single stream DiT with 32 layers, Qwen3-VL-8B as the text encoder, a new VAE with 64 channels and RGBA support, flow matching. Native 2K carried over from 2.0.

So the image tower went from 20B to 7B while the text encoder grew. The model got better at reading your prompt and cheaper to run at the same time, which is why a machine that could not hold the 2025 model comfortably can hold this one.

What 2.1 actually added

Four things, none of them cosmetic:

1. A native alpha channel. RGBA straight out of the VAE, so a cutout is a property of the model rather than a second tool.
2. Up to ten reference images in one pass. Compose from several inputs instead of one.
3. Local edits from a drawn marking, not only from a mask.
4. A prefix KV cache, which is what makes repeated edits on the same image quicker, plus better typography than any Qwen image model before it.

There is no Lightning or Turbo LoRA for 2.1 as of 21 September 2026, and the existing ones do not carry over. They were built for a different architecture and will not load.

The licence, without the sugar

Qwen-Image 2.1 ships under the Qwen Research License, not Apache 2.0. Research and evaluation only. No commercial use without a separate licence from Qwen. That is the whole of it, and it is worth reading twice if you were planning to put this model behind a product.

The older models are still there and still permissive: Qwen-Image 2512 and Qwen-Image-Edit 2511 remain Apache 2.0. If your work is commercial, those two are the ones you want, and nothing about 2.1 changes their status. We would rather say this plainly on the way in than have somebody find out on the way out.

What it needs from your machine

The ComfyUI repack from Comfy-Org ships three files, and the int8 build is the one most people will run:

File	Size as Hugging Face shows it	On disk
qwen_image_2.1_int8_convrot.safetensors	7.26 GB	6.76 GiB
qwen3vl_8b_int8_convrot.safetensors	9.35 GB	8.71 GiB
qwen_image_2.1_vae_bf16.safetensors	676 MB	0.63 GiB

The full bf16 versions are 14.2 GB for the image tower and 17.5 GB for the encoder. A w4a8 encoder build also exists at 6.31 GB, and it is not the one used here.

For memory, the figures below come from the ai.rs write-up rather than from our own bench, and should be read as a guide rather than a measurement:

• The int8 pipeline lands at roughly 16 GiB in total, split 6.8 for the image tower, 8.7 for the text encoder and 0.6 for the VAE.
• A 24 GB card holds the whole thing resident, which is the comfortable case.
• On a 16 GB card, move the text encoder into system RAM and the rest fits.
• Full bf16 wants 33 to 40 GB, so that is workstation territory.

The same source reports 5.93 seconds for a single image on an RTX 5090. That is their number on their hardware, not ours. We have not produced timings of our own, and a second hand figure is worth exactly what it is worth.

If none of those tiers describes your machine, our guide to generating AI images on your own computer works through what each VRAM class can actually hold.

Running it in LU Labs 3.0.1

LU Labs is the free desktop studio for Windows and Linux, and version 3.0.1 carries Qwen-Image 2.1 as a Model Manager bundle. There is no Mac build.

Install. Open the Model Manager and look for Qwen-Image 2.1 (Generate and Edit): three files, 16.1 GB, with "Qwen Research License, non-commercial use" written into the bundle description so nobody has to go hunting for it. One click puts all three into the right ComfyUI folders. The app marks the bundle as installed as soon as the running ComfyUI lists the files, so if you already downloaded them by hand they are recognised rather than fetched twice.

The version gate. The model needs ComfyUI 0.37.0 or newer, because the TextEncodeQwenImage21 node landed in that release and the older edit nodes do not fit 2.1. If your backend is behind, the app says so in as many words: "Qwen-Image 2.1 needs ComfyUI 0.37.0 or newer. Update ComfyUI in Settings." The update itself is under Settings, AI Backends, ComfyUI.

Generating. After the install, 2.1 appears in the model picker of both the Image lane and the Edit lane in the Create tab, like any other local image model. No new button, no new tab. In the Image lane you write a prompt and, if you want one, a negative prompt. The defaults follow Qwen's own template: 25 steps, CFG 1, euler, simple, 1024x1024. CFG below 1 is not usable with this model, which is worth knowing before you go exploring. Advanced settings are the usual ones: steps, CFG, width, height, seed and batch.

Editing. The Edit lane takes one reference image plus a prompt, and no mask. The mask route is deliberately closed for this model in 3.0.1, and the output follows the reference image. Resolution follows the canvas you pick. If you are used to the masked edit workflow from other models, and our Flux comparison covers who owns that lane, this is a different shape of tool: you describe the change instead of painting where it goes.

The LoRA stack is open for every local image model in the app, Qwen-Image 2.1 included. There is simply nothing to load into it yet.

What is not in this release

Being clear about the gap between the model card and the build:

• Multiple reference images. The model takes up to ten, the 3.0.1 Edit lane takes one. Not in this release.
• RGBA output. The VAE can do it, the app does not expose it. Not in this release.
• Masks with this model. Not in this release.
• A prompt enhancer and a KV cache node. Neither is in this release.

Everything above is a description of the build, not a plan. When those things exist, there will be a post about them.

One more piece of honesty about how this was checked. Version 3.0.1 shipped with unit tests and a visual pass over the interface. Nobody on our side generated an image with these weights and timed it, so you will not find a quality or speed claim from us anywhere in this post. The numbers here are Qwen's, Hugging Face's and ai.rs's, each one labelled.

Getting started

Download LU Labs, free, Windows and Linux. Install ComfyUI from Settings, AI Backends if you do not have it, make sure it is on 0.37.0 or newer, then take the Qwen-Image 2.1 bundle from the Model Manager. First generation, defaults, 1024x1024, and go from there.

And then the boring question that decides everything else: is your work commercial? If it is, stay on Qwen-Image 2512 and Qwen-Image-Edit 2511, which are Apache 2.0 and still very good. If it is research or evaluation, 2.1 is the better model, and it is one download instead of two.