Lightricks

/

LTX-2.5

like 1.88k

Follow

LTX.io 4.8k

I made one for the text encoder.

https://civitai.com/models/2860946/gemma4-12b-with-proj-ltx-25-w4a8-under-10gb

I made one for the text encoder.

https://civitai.com/models/2860946/gemma4-12b-with-proj-ltx-25-w4a8-under-10gb

Thanks, but compared to the TE, what I need more is the Diffusion Model’s unpruned_w4a8_mixed.safetensors weights.

I did not try to make that yet. I doubt it will be smaller than the nvfp4.

have you tried this https://huggingface.co/Winnougan/ltx-2.5-w4a8-convrot-int4-convrot-Winnougan-Blessing/tree/main/diffusion_models

Mine is a bit bigger but in early testing pretty close to int8 https://civitai.com/models/2867691/ltx-25-distilled-w4a8?modelVersionId=3239819

Mine is a bit bigger but in early testing pretty close to int8 https://civitai.com/models/2867691/ltx-25-distilled-w4a8?modelVersionId=3239819

https://huggingface.co/tsolful/LTX_2.5_INT4_W4A8_ConvRot

https://huggingface.co/tsolful/LTX_2.5_INT4_W4A8_ConvRot/discussions/2

As stated by the author, this model weight is consistent with Kijai’s W4A8‑mixed quantization.