Qwen-Image-2.1 runs on one consumer GPU
Qwen-Image-2.1 cuts the visual generator from 20B to 7B. BF16 weights are 14.2 GB, INT8 is 7.26 GB, and a Q8 run sits near 16 GiB, up to 17 GB at 50 steps.
Alibaba released Qwen-Image-2.1 on 20 September 2026. The visual generator is 7B parameters, down from 20B in the prior Qwen-Image. BF16 weights are 14.2 GB, down from 40.9 GB. An INT8 build is 7.26 GB. It runs on consumer GPUs with 24GB of VRAM or less, especially quantized.
The weight cut and the measured runtimes are in eesel's Qwen Image 2.1 report. This page is the hardware and speed record: weight size, VRAM, and seconds per image. It is not native RGBA, not the Qwen Research License, which bars commercial use without a separate agreement, and not editing from up to ten reference images.
The generator file got smaller
BF16 here is the weight size reported for the visual generator. INT8 is the smaller build of that generator. The prior model has no INT8 figure in the comparison.
| Metric | Qwen-Image (prior) | Qwen-Image-2.1 |
|---|---|---|
| Parameters | 20B | 7B |
| BF16 | 40.9 GB | 14.2 GB |
| INT8 | not listed | 7.26 GB |
7B is a bit over a third of 20B. 14.2 GB is a bit over a third of 40.9 GB. The review of the same release repeats the cut in one line: a 20B model with a 40.9 GB BF16 image transformer, then a 7B visual generator at 14.2 GB BF16 and 7.26 GB INT8.
The 7B count is only the visual generator: a 32-layer single-stream diffusion transformer. The text encoder is a separate Qwen3-VL 8B. The VAE is a 64-channel RGBA VAE with 16x spatial compression. Apidog's launch notes list those three pieces side by side. This page does not add 7B and 8B into a new total. The published comparison is 20B to 7B, and 40.9 GB to 14.2 GB, for the generator.
One checkpoint does text-to-image and image editing, so those jobs are not two weight files. What a transparent layer can do, and how ten references are composed, is on the other pages.
What the GPU actually holds
The 7.26 GB INT8 file is a weight size. It is not the memory a running job holds. Q8 is the label on the runtime figure: about 16 GiB. A 50-step run goes up to 17 GB. Coverage prints those units as they stand. This page does not convert GiB into GB, and it does not fold the two figures into one range.
CPU offloading moves work off the GPU. It can cut VRAM to ~3 GB. That figure is the offload case. It is not the Q8 figure, and it is not the 17 GB peak of a 50-step run.
Tech Insider's account of the 20 September release puts the model on consumer GPUs with 24GB of VRAM or less, especially quantized. 24GB is the class named for a quantized run. Reserve the live job: about 16 GiB at Q8, up to 17 GB across 50 sampling steps, or ~3 GB with CPU offloading. The text encoder is still the separate Qwen3-VL 8B, and the VAE is still a separate piece, so the INT8 filename is not the whole load.
| Runtime figure | Reported value |
|---|---|
| VRAM at Q8 | about 16 GiB |
| VRAM during a 50-step run | up to 17 GB |
| VRAM with CPU offloading | ~3 GB |
Seconds per image, on named cards
Generation time is how long the job takes to finish. The reported band for a 1MP image is 5-25 seconds on modern GPUs. "Modern GPUs" does not name the card at either end of that band. The band is the report for a 1MP still.
| Job | Time | Where it was measured |
|---|---|---|
| 1MP image | 5-25 seconds | modern GPUs |
| 4K render | ~90 seconds | RTX 5080 |
| 2K edit at 25 steps | ~2 minutes | RTX 5060 Ti, community testing |
The 4K row is one card and one job: ~90 seconds on an RTX 5080. The edit row is community testing. A 2K edit at 25 steps took about 2 minutes on an RTX 5060 Ti. The eesel report is where that test is written down.
A 25-step edit and a 50-step run are not one experiment. The 17 GB peak is tied to 50 steps. The ~2 minute edit is tied to 25 steps at 2K on that 5060 Ti. No published measurement joins those two rows, so neither row is a formula for the other.
DataNorth's release note records the architecture and the day-one tools. The seconds above are the reported runs.
Same-day weights, day-zero loaders
Weights landed the same day on Hugging Face, ModelScope, and GitHub. Day-zero support shipped in ComfyUI, Diffusers, vLLM-Omni, and SGLang. A local run did not have to wait on a loader written after the announcement. Those names mean the weights were runnable on release day. They are not a score against a closed model.
A finished run is not a win
No reproducible numerical benchmark table against closed rivals was published in the initial coverage wave. Tech Insider says the early write-ups do not contain a head-to-head numerical table against closed models. A generator that fits a consumer GPU can still lose the picture. "It runs" is not "it wins."
Community testers scored Qwen-Image-2.1 at 7/15, against 4/15 for the prior model. That is a real improvement on the previous Qwen-Image, with visible synthetic-data artefacts. It is a score against that prior model. It is not a score against a closed rival, and it is not the missing table.
Reviewers describe a yellowish colour tone and graininess, read as heavy GPT-Image training-data influence. The model struggles with infographics and adds unsolicited Chinese text unprompted. Text rendering is contested: some call it much better, others still rate Ideogram superior for clean small text. On complex compositional requests, the text encoder gets overloaded with larger prompts.
A 1MP image can return in 5-25 seconds and still be yellow, grainy, or lettered wrong. The ~90 second 4K render on an RTX 5080 is a timing. The ~2 minute 2K edit on an RTX 5060 Ti is community testing. Speed says whether the card finishes. The 7/15 score, the colour cast, and the absent closed-rival table say whether the file is the one to keep.