LTX-2.3

What it is

LTX-2.3, the video model from Lightricks, is the cleanest legal proposition in this section. It is released under the Apache 2.0 licence, free to use for any organisation under $10m in annual recurring revenue, and, most usefully, it is trained on licensed data. That combination gives LTX-2.3 the strongest legal position of any model covered here, and it is the reason the model deserves attention from anyone who has to think about where their training data came from.

The licence

Start with the licence, because it does a lot of work. Apache 2.0 is a permissive open-source licence: it lets a business use, modify and build on the model commercially, without the copyleft obligations that force you to open-source your own work in return. Lightricks pairs that with a revenue threshold, so a company earning under $10m a year in recurring revenue pays nothing. For a startup or an independent studio, that removes the model from the list of things to budget for, and it removes the negotiation that a closed API always eventually requires.

Training data

The training-data point is the one that will decide procurement conversations. Most video models are trained on data of uncertain provenance, scraped from the open web, and that uncertainty becomes a liability the moment a client, a broadcaster or a regulator asks the awkward question about rights. LTX-2.3 is trained on licensed data, which means Lightricks can stand behind what went into it. For any business that has to answer a procurement question about training-data provenance, this is the page to send them to. Legal comfort is rarely a glamorous feature, yet it is often the deciding one when a model moves from a creative experiment to something a company ships in a product.

Length and audio

On capability, the standout figure is length. LTX-2.3 produces the longest single-pass clips in the category, at around 20 seconds. Single-pass is worth understanding in the technical sense: the clip is generated as one continuous piece rather than assembled from shorter segments stitched together, which tends to hold motion and continuity more convincingly across the length of a shot. Most rivals stop well short of 20 seconds in a single generation, so a longer coherent take is a genuine practical advantage for anyone who needs a shot to run rather than cut. The model also added stereo audio at 24kHz in March 2026, so a generation now carries a soundtrack with left-right separation rather than a flat mono bed.

Where it sits

Lightricks is a consumer software company that also ships editing apps, and that heritage shows in how the model is packaged for people who cut video for a living. The code is available on the company’s public repository, so it can be run and inspected rather than accessed only through a metered endpoint, which reinforces the open-source claim. A company that has spent years building tools for editors tends to think about outputs in terms of a working timeline rather than a single showpiece clip, and that shows in the emphasis on longer takes and usable audio. The revenue threshold works commercially too: it lets Lightricks seed the model widely among smaller users while reserving paid terms for the larger firms that can afford them, a distribution approach the open-source world has used to good effect before.

Where LTX-2.3 sits in the wider AI video field is clear enough: it is the model to reach for when the legal position weighs as heavily as the picture. Alongside a small number of other genuinely open options, it forms the part of the market that competes on licence terms and data provenance as much as on raw quality. As generated video moves from novelty into commercial production, that clean legal footing is likely to count for more, not less, and Lightricks has staked out that ground earlier and more deliberately than most. For a broader view of how the open models compare, see the AI hub.