Qwen Image 2512: What Changed and How to Run It Locally (GGUF, Mac)

Qwen-Image-2512 is the December 2025 update of Qwen-Image, the 20B open text-to-image model from Alibaba's Qwen team. It is released under Apache-2.0, so you can use the images commercially. This guide covers what changed, which files to download, what hardware it needs, and how it compares with the newer Qwen-Image 2.1.

What is Qwen-Image-2512?

Qwen-Image-2512 is a new checkpoint of the original Qwen-Image text-to-image model released in August 2025. The name is the date: 2025, December. The Qwen team lists three improvements:

• More realistic people. Less of the "AI-generated" look, especially in faces and skin.
• Finer natural detail. Landscapes, animal fur and other natural textures are more detailed.
• Better text rendering. More accurate lettering and layout, in both Chinese and English.

The team reports that it was the strongest open-source model in more than 10,000 rounds of blind comparison on its AI Arena leaderboard at release.

It is a text-to-image model only. For editing, the matching models are Qwen-Image-Edit-2511 and 2509.

Qwen-Image-2512 vs Qwen-Image 2.1

These are two different model lines, and the choice matters mostly for license and hardware:

Pick 2512 if you need commercial use or already have a Qwen-Image workflow. Pick 2.1 if you have less memory, want one model for generation and editing, or want the official 8-step Turbo. We tested 2.1 Turbo on a Mac in our local Qwen-Image 2.1 Turbo test.

Files and hardware

The official repository is large because the weights are in BF16:

That is too much for most consumer GPUs. A quantized GGUF is the practical way to run it locally.

Which Qwen-Image-2512 GGUF to download

The unsloth/Qwen-Image-2512-GGUF repository has the full range:

These sizes are the diffusion model only. Add the text encoder (a quantized one is a few GB) and the VAE. With offloading to system RAM, a smaller GPU can still finish an image, but it will be slower.

How to run Qwen-Image-2512 locally

ComfyUI with GGUF

1. Update ComfyUI and install the ComfyUI-GGUF custom node.
2. Put the GGUF file in models/unet/ and load it with the Unet Loader (GGUF) node.
3. Use the Qwen2.5-VL text encoder and the Qwen-Image VAE named by your workflow template.
4. Start at the official settings: about 1328 × 1328 for a square image, 50 steps, true CFG 4.

Fewer steps

Fifty steps is slow. Two community options cut it down:

• lightx2v/Qwen-Image-2512-Lightning, a distilled LoRA for a few steps.
• Wuli-art/Qwen-Image-2512-Turbo-LoRA, with a 2-step variant.

Use the step count and CFG from the LoRA's own card. Settings from the base model do not transfer.

On a Mac

Community MLX conversions exist, such as machiabeli/Qwen-Image-2512-4bit-MLX, alongside 3-, 5-, 6- and 8-bit versions. A 20B model plus its encoder needs a lot of unified memory, so plan on 32 GB or more, even at 4 bits.

Tips for better results

• Write the text in quotes. Keep each piece of lettering short, and say where it goes.
• Describe the photo, not the quality. "Casual phone snapshot, indoor light" works better than "masterpiece, 8k".
• Use the official aspect ratios. For example 1664 × 928 for 16:9 or 928 × 1664 for 9:16.

If you want a one-click app instead

TopLocal Studio does not ship Qwen-Image-2512. Its Qwen option is the newer Qwen-Image 2.1 Turbo, next to Z-Image-Turbo and FLUX.2 klein. It runs on Mac and Windows with no setup, and everything stays offline. See the homepage for details.

FAQ

Is Qwen-Image-2512 free for commercial use?

Yes. It is released under Apache-2.0. Community GGUF and LoRA files keep the original model's license.

What is the difference between Qwen-Image and Qwen-Image-2512?

2512 is the December 2025 update of the same 20B model. It gives more realistic people, finer natural detail and better text.

How much VRAM does Qwen-Image-2512 need?

The full BF16 model is 40.9 GB. With a Q4_K_M GGUF (13.2 GB), a 16 GB GPU is the comfortable minimum; smaller GPUs need offloading and run slower.

Can Qwen-Image-2512 edit images?

No, it is text to image only. Use Qwen-Image-Edit-2511 for editing, or Qwen-Image 2.1, which does both.

Is Qwen-Image-2512 better than Qwen-Image 2.1?

They target different needs. 2512 is larger and can be used commercially. 2.1 is smaller, newer, edits as well, and has an official 8-step Turbo, but its license is non-commercial.