More qwen_image_2512 quants

#18

by DocterZed - opened Jan 14

Discussion

DocterZed

Jan 14

Please add more qwen_image_2512 quants, especially the fp8mixed or fp8_hq.

Thanks.

kajaks

Jan 16

Also, nvfp4 would be great - the present one is 2508, judging from #17

alcoartist

Jan 17

So qwen_image_fp8_hq.safetensors is not 25.12 version despite the publishing date?

kajaks

Jan 20

Barring comment from one of the maintainers, I strongly assume qwen_image_fp8_hq.safetensors is not based on Qwen Image 2512, despite the recent
upload date.

I let Claude compare the actual tensor values (e.g., img_in.bias) across files.
The fp8_hq weights are byte-identical to the old (2508) model, not to 2512.
The same applies to nvfp4 and fp8mixed – all three are quantizations of the old 2508 model.

I gather that only qwen_image_2512_bf16.safetensors and qwen_image_2512_fp8_e4m3fn.safetensors
are actual 2512 models.

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment