LTX 2 open source: what Lightricks released, and the license
LTX-2 is an audio-video model with open weights under the LTX-2 community license. What the Hugging Face cards list, and why Sume's catalog has no LTX.
Yes, LTX-2 has open weights: Lightricks publishes the checkpoints on Hugging Face, and the card describes it as a DiT-based audio-video model that generates synchronized video and audio in one model. It is released under the ltx-2-community-license-agreement, not a standard Apache or MIT license, so read that agreement before commercial use.
Facts about LTX are from the LTX-2 card and the newer LTX-2.3 card, read 2026-09-29. Sume facts are from Video generation.
What is in the LTX-2 release?
The LTX-2 card lists several checkpoints. The card also says a newer version, LTX-2.3, is available, and describes it as a significant update with improved audio and visual quality and enhanced prompt adherence.
| Checkpoint | Card note |
|---|---|
ltx-2-19b-dev | The full model, flexible and trainable in bf16 |
ltx-2-19b-dev-fp8 | The full model in fp8 quantization |
ltx-2-19b-dev-fp4 | The full model in nvfp4 quantization |
ltx-2-19b-distilled | The distilled version of the full model, 8 steps, CFG=1 |
ltx-2-spatial-upscaler-x2-1.0 | An x2 spatial upscaler for the ltx-2 latents, for higher resolution |
What does it take to run LTX-2 locally?
The card says the codebase was tested with Python 3.12 or newer, a CUDA version above 12.7, and supports PyTorch ~= 2.7. It lists ComfyUI nodes and a Diffusers pipeline as ways to run it. It gives no minimum VRAM figure, so this post gives none.
One input rule from the card: width and height must be divisible by 32, and the frame count must be divisible by 8, plus 1. Otherwise pad the input and crop the output.
What is LTX-2.3?
The LTX-2.3 card describes LTX-2.3 as a significant update to LTX-2 with improved audio and visual quality and enhanced prompt adherence. Its checkpoint names carry a 22b size (for example ltx-2.3-22b-dev and ltx-2.3-22b-distilled), where LTX-2 uses 19b. It has the same ltx-2-community-license-agreement label.
The card says Diffusers support for LTX-2.3 is coming soon, so use the card's own code for it today.
What are LTX-2's stated limits?
- It is not intended or able to provide factual information.
- It may fail to generate videos that match the prompt perfectly, and prompt following depends heavily on prompting style.
- When generating audio without speech, the audio may be of lower quality.
Is LTX available through the Sume API?
No. Sume's catalog code lists these video ids: seedance-2.5, seedance-2-mini, seedance-2, seedance-2-fast, kling-3, wan-3.0, grok-imagine-video-1.5, minimax-h3, minimax-h3-max, gemini-omni-flash-1.1. No LTX model is among them and the docs do not mention one. If you need sound generated with the clip, the models that report an audio capability are in the catalog's generate_audio field; Does the video model make the audio, always or optionally? explains it.
For the run-it-yourself question, see Open-source video model vs API.
Sources
Related posts
More in Models
- Lyria 3 Pro or Lyria 3.5: which model id do I send?
Sume's Music Router accepts sume/music-auto, lyria-3.5 and lyria-3-pro. What each id does, how to pin one, and what Google lists for the models.
- Lyria 3.5 API: model id, request and how to call it
Lyria 3.5 is in the Gemini API as lyria-3.5. Here is what Google's page says about it and how to call it from Sume with one music request.
- AI music generator with vocals: Lyria 3.5 lyrics by API
Lyria 3.5 can sing. How to ask for vocals or an instrumental, steer lyrics with section tags, and read the lyrics back from a Sume music job.
- Lyria 3.5 release date: where it is available
Google announced Lyria 3.5 in the Gemini app on 2026-09-04. Where Google says you can use it, and where Sume lists it.
Written by Sume