On March 5, 2026, Lightricks — the company best known for Facetune and its LTX Studio product — did something most well-funded AI labs still won’t: it open-sourced the full weights, training code, and architecture for a frontier-class video generation model. LTX-2.3 is a 22-billion-parameter diffusion transformer, and unlike almost everything else in its weight class, it doesn’t just generate picture. It generates picture and sound in the same forward pass.
Watch: LTX-2 Unleashed AI Video (YouTube)
What actually shipped
LTX-2.3 splits into two coordinated streams — a 14-billion-parameter video stream and a 5-billion-parameter audio stream, linked by bidirectional cross-attention so each one can influence the other while generating. The practical result: native 4K output at up to 50 frames per second, with synchronized stereo audio, in clips up to 20 seconds, in both horizontal and 9:16 portrait framing. It supports text-to-video, image-to-video, video-to-video, audio-to-video, and depth conditioning, and — critically for anyone without a render farm — it’s built to run on consumer GPU hardware rather than requiring datacenter-scale compute.
The weights and code are published on GitHub and Hugging Face under commercial-friendly terms: no subscription, no usage cap, self-hosting explicitly allowed. That puts it in a different category from Sora, Runway, or Kling, all of which keep their best models gated behind a paid API.
Why the audio part is the actual story
Plenty of open video models have shipped this year with better raw image fidelity. What’s made LTX-2.3 the one people keep bringing up isn’t resolution — it’s that sound stopped being an afterthought. Most generative video pipelines still treat audio as a separate step: generate the picture, then hand it to a different tool (or a human) to score, foley, or dub. A model that produces both together, with the two streams actually informed by each other during generation, removes an entire pipeline stage that independent creators have historically had to work around by hand.
That’s the detail worth paying attention to if you’re evaluating whether to build on it: the value isn’t “better video,” it’s “one fewer tool in the chain.”
Where to start
Because the weights are fully open, the most realistic entry point for most readers is the official documentation rather than a hosted demo — running it locally means dealing with real VRAM requirements, but also means no per-generation billing once you’re set up. If you already have a ComfyUI or similar node-based pipeline, an open-weight model in this class is the kind of release that tends to get community-built integrations quickly; check current node repositories before assuming you have to write your own loader from scratch.