Lightricks

Five students in Jerusalem, and an app that reached hundreds of millions of devices

Lightricks was founded in February 2013 on the Hebrew University campus in Jerusalem by five PhD students and graduates: Zeev Farbman, Nir Pochter, Yaron Inger, Amit Goldstein and Itai Tsidon. The first product, Facetune, for editing portraits, came out that same year and was chosen as one of the best apps of the year in Apple's App Store. It was followed by more photo and video editors, including Videoleap for video editing, which came out in September 2017 and was chosen as Apple's app of the year. The business model was clear: a consumer app on a monthly subscription, with very low running costs, because the filter or fix runs on the phone's own processor and not on a company server. By 2021 the company reported more than half a billion downloads of its apps and more than five million paying subscribers. Revenue rose from about $17 million in 2017 to about $52 million in 2018, with a profit of about $10 million. This model is also what allowed a company from Jerusalem to reach a global market without a sales force: users discover the app in the store, download it and pay for it themselves, and distribution is handled by Apple and Google.

Funding and valuation

In July 2019 Lightricks raised $135 million in a Series C led by Goldman Sachs, at a valuation of $1 billion, becoming a unicorn. In September 2021 it raised about $100 million more in a Series D, at a valuation of about $1.8 billion. Here again it is worth telling apart two numbers that are easy to confuse: $1.8 billion is the valuation investors agreed on in the deal, not a sum that went into the company's bank account. The amount that actually came in during that round was much smaller. What makes Lightricks unusual in the Israeli tech landscape is that it built a large business on individual consumers rather than organizations. That is a relatively rare path here, in which revenue comes from millions of users who each pay a little, instead of from hundreds of business customers who pay a lot. This choice sounds better (millions of small users instead of dependence on a few huge customers), but it is also vulnerable in its own ways. It depends on the app stores, the fees they charge and their policy changes, and also on users remembering to keep paying month after month.

The turning point: from editing existing content to creating new content

In 2024 the company changed direction. In February it announced LTX Studio, an environment that produces a whole video from a written script, and it opened to the public in August; in November of that year it released the video model LTX-Video. The difference between these two and what came before is fundamental, not gradual. A classic editing app gives users tools to change content they have already shot: cut, brighten, blur. A generative model produces content that did not exist at all, from a text description. This difference shows up mainly in the electricity bill. A filter runs on the phone itself and costs the company almost nothing; generating one second of video requires an expensive compute accelerator in the cloud, and costs real money every single time. A low monthly subscription business model, which fits the first case perfectly, breaks down completely in the second. A change like this is not only a matter of cost but of organizational skill. A company that knew how to build a fast, pleasant phone interface suddenly had to train models on computing clusters, manage datasets at huge scale, and compete with research labs that were much bigger and better funded.

Open weights: what they are, and why this is an unusual choice

Lightricks's unusual choice was to release its video models with open weights. Weights are the numerical parameters learned in training: in effect, the model itself. When they are released, anyone can download the file and run it on their own hardware, instead of using the model only through the company's online service. On 23 October 2025 the company introduced LTX-2, which according to the company generates video and sound together in the same process, at 4K resolution and up to 50 frames per second, and runs efficiently on consumer graphics cards. The release was done in stages: the announcement published the core components, the datasets and the inference tools, and the weights themselves were promised for a later date and released in January 2026 together with the rest of the code. More versions followed, including LTX-2.3 in March 2026, which came with a desktop video editor that runs the entire model locally. The company kept up this pace afterward too, and in August 2026 it introduced LTX-2.5, which according to reports generates a ten-second video in less than seven seconds.

Why would a company give away a model that cost it a lot to develop

The answer is not only ideological; it has clear business logic. If the model runs at the user's or the business customer's site, on their hardware, the running cost passes to them, and the company can sell a complete work environment, service and tools instead of selling expensive computer time that it pays for itself. An open release also builds a developer community around the model, produces improvements and add-ons the company would not have built itself, and sets a standard that a closed competitor finds hard to compete with. On the other hand, the price is clear: whoever downloaded the model does not pay for it, and the company willingly gives up direct revenue from its most valuable asset. This is one of the few times an Israeli company has published a foundation model for the community instead of keeping it closed, and in doing so Lightricks became a live test case for the question of whether a sustainable business can be built around a model anyone can download. This debate, closing a model or opening it, is one of the central debates in the whole AI industry, and Lightricks is one of the few Israeli examples on the open side of it.

2025-2026: layoffs and a split into two entities

The answer to that question, as of today, is not clear-cut. The cost of developing and running video models turned out to be a heavy burden, and the company carried out two rounds of layoffs of about 15% each, in October 2025 and in June 2026, together affecting about 160 employees. In June 2026 the company also announced a split into two separate entities: one around the generative video models, and the other around the long-standing editing apps that continue to bring in revenue. This split is essentially an admission that these are two very different businesses: one with millions of paying users and negligible running costs, and the other with high running costs and a revenue model that is still taking shape. Whether a consumer product company can carry the development of a foundation model on its own is exactly the question this case tests, and the provisional conclusion, at least in organizational terms, is that it is hard to do under one roof. As of today, both entities continue to operate, and the models keep coming out in new versions. The question here is not whether the technology is good (it is) but who ends up paying for it: the user, the business customer, or the investor.