Header too large

[VIPQualityPost]

Trying to run this, I can't seem to even load the model:
safetensors_rust.SafetensorError: Error while deserializing header: header too large
I have a 24GB 3090, from the readme it seems like this should be enough to run it?
I'm on Ubuntu 25.10.

[maybleMyers]

where did you get the model from?

[VIPQualityPost]

I got it from the links in the upstream project README (hosted by LTX on hf).
Do you think the model is corrupted? I tried both with 19b-dev and 19b-dev-fp8 (with fp8 setting enabled).

[maybleMyers]

It might be incomplete or corrupted. I just use the ones from the main ltx repo. It is not a vram problem. It would give you a cuda memory error.

The 43GB file didn't get fully download perhaps. Check file size:
ls -lh weights/ltx-2-19b-dev.safetensors
Should be ~43GB for dev, ~27GB for fp8
Using huggingface-cli (recommended)
pip install huggingface-hub
huggingface-cli download Lightricks/LTX-2 ltx-2-19b-dev.safetensors --local-dir ./weights
or
hf download Lightricks/LTX-2 ltx-2-19b-dev.safetensors --local-dir ./weights