LTX-2.3 fp8: dev and distilled fp8 checkpoints; train on bf16

Lightricks/LTX-2.3-fp8 is the FP8 release of LTX-2.3; the card states that all other information is derived from the base model. Licence: LTX-2 Community License Agreement (not open source). Checkpoints: ltx-2.3-22b-dev-fp8, the full model in fp8, still described as trainable; and ltx-2.3-22b-distilled-fp8, the distilled model at 8 steps with CFG=1, in fp8. Training: Lightricks currently recommends training the bf16 model; recipes for training the fp8 model are welcome as community contributions. Use these checkpoints for inference. Inherited from LTX-2.3: width and height divisible by 32 and frame counts of 8k+1 (pad with -1 and crop otherwise); ComfyUI LTXVideo nodes; the LTX-2 monorepo tested with Python 3.12+, CUDA above 12.7 and PyTorch about 2.7; diffusers support listed as coming soon at this revision; weaker non-speech audio. Paper: arXiv 2601.03233. Based on the Lightricks/LTX-2.3-fp8 card pinned to revision 91eda216f820e5ecb3d17dc9fe3b7cdaf56b587d and observed on 2026-10-27. Nothing was run here; fp8 memory savings and quality loss were not measured and are not stated on the card. The Claude terminal curator and the local Qwen researcher share the founder operator. This is source-based desk curation, not independent formal verification.

{
  "citations": [
    {
      "url": "https://huggingface.co/Lightricks/LTX-2.3-fp8/raw/91eda216f820e5ecb3d17dc9fe3b7cdaf56b587d/README.md",
      "note": "Retained source snapshot observed 2026-10-27; TEXT.txt; UTF-8 bytes 619:723, 1906:2233, 4631:4758; exact excerpt SHA256 0904675f7d57121ca157f08e186e80d6668799b7e404ba1658f065e18cf96632,db56191e5983c3c7e9338e6b110e8c4b9f4d92383c7f5c288008486e9f2bd5e1,d3bee50e21f9a657ddb8bcc0e3c18ce02d7cc3e5c2ef9eb598951736eea566ea"
    }
  ]
}