How long a clip the published weights can make
The LTX-Video repository states support for up to 60 seconds of video on its 13B model, with distilled and fp8 variants published for lower memory and faster inference. Two other repositories here also reach sixty seconds, each writing it in a different unit. As of 2026-09-12.
| Item | What is published | Detail |
|---|---|---|
| LTX-Video 13B | Up to 60 seconds | Stated in the repository |
| Published variants | Distilled and fp8 | For lower VRAM and faster inference |
| Hosted single generations | Tens of seconds at most | Much shorter than the open figure |
Inclusion rule. Lengths stated in the repository. A length achieved by chaining generations is not the same claim and is not recorded. Order. From the open figure to the hosted comparison.
1A longer ceiling locally than commercially is worth pausing on
Hosted services sell short generations, and the usual explanation is model capability. This row suggests the constraint is at least partly commercial: an open release of comparable size publishes a much longer figure.
The register records the published numbers and does not assert the explanation. What it can say is that the shorter hosted ceilings are not obviously a statement about what the architecture can do.
2Sixty seconds is a supported length, not a promise about quality
Support for a duration means the code path exists and runs. Whether a minute of generated video holds together is a different question, and one nobody publishes an answer to in either direction.
A production reading this should treat it as a capability to test rather than a result to rely on, which is how every published figure in this register should be read.
3Variants are the vendor telling you the hardware story
Publishing distilled and fp8 builds says the vendor expects a range of hardware and wants the model reachable across it. That is a clearer signal than any minimum specification, because it is an action rather than a claim.
It also complicates the length figure: nothing says the sixty-second ceiling holds on the smaller variants, and the repository does not break the figure down by build.
- LengthThe repository states support for up to 60 seconds of video on LTXV-13Bfrom the release notes
- Published modelsA 13B model, with distilled and fp8 variants published for lower VRAM and faster inferenceseveral variants
4Open weights, hosted
Instead of local hardware for “Clip length”, SceneMixer runs hosted models; its guides describe the result:
AI dubbing replaces the audio of a finished clip and then tries to match the lips. Native-language generation gives the video model the line in the target language before the shot exists, so mouth movement, pauses and shot length are produced for that language.
SceneMixer, Native-Language Dialogue vs. Dubbing in AI Video (2026): Why the Cast Should Speak the Line — “Is native-language generation the same as AI dubbing?” (read 2026-10-13)
SceneMixer writes dialogue into the script analysis, assigns each line to a shot with a speaker and a short delivery note, and the video model performs it — picture and audio generated together, so the mouth forms that sentence in the language you chose.
SceneMixer, AI Dialogue Generator (read 2026-10-13)
5Sources
Repository statements and hosted rates are linked from the model pages and the method page, checked 2026-09-12. Hardware figures are set out on requirements. Nearby: Buy or run, Same licence, different reach, Floor or peak.