LTX-2 is an open-source video generation model from Lightricks, the company behind Facetune and LTX Studio. It builds picture and soundtrack in a single pass, so dialogue, sound effects and ambience land in sync with the action on screen. Announced in October 2025, it saw its full weights published a few weeks later, free to download. Output reaches native 4K at 50 frames per second, either locally on a consumer GPU or through an API billed by the second.
LTX-2 is built on a two-stream diffusion model, 14 billion parameters for video and 5 for audio, tied together by cross-attention layers. In practice, a door slamming on screen also slams in the soundtrack, and lips follow the dialogue with no manual fixing.
You type a description or drop in a reference image, pick a length and a resolution, and the clip arrives with its mix baked in. Version 2.3 tightened things further: sharper detail, steadier motion, vertical format supported.
Everything required to run the model without the cloud is public, a level of openness still rare among open-source AI projects. The community license covers commercial work as long as your company stays under $10M in annual revenue; above that, a dedicated license applies.
The published package includes:
On the hardware side, plan for an NVIDIA card with 32GB of VRAM to be comfortable on Windows, plus roughly 150GB of model files on first launch (brew a coffee, this is no five-minute download). After that, nothing to pay per clip.
Against Veo 3.1, Seedance 2.0 or Kling 3.0, all closed and billed by the credit, LTX-2 plays the control card: a pipeline you can inspect, fine-tune and deploy on your own infrastructure. Published comparisons put it ahead for high-volume work and technical teams, while proprietary rivals often keep an edge on pure photorealism.
Four routes lead to the model, from tinkerer-grade to armchair-comfortable. Terms shift with each release, so treat the official ltx.io page as the only reference worth checking before spending anything.
| Access route | Hardware required | Billing |
|---|---|---|
| Open weights (Hugging Face, ComfyUI) | Beefy NVIDIA GPU, manual setup | Nothing per clip |
| LTX Desktop | Windows or Mac, 32GB VRAM recommended | Nothing when local |
| API (Fal, Replicate, LTX) | None, everything runs in the cloud | Per second generated |
| LTX Studio | Any browser | Credits and subscription |
Yes, the weights download freely and local generation costs nothing per clip. The community license allows commercial work up to $10M in annual revenue. Only the API, billed per second of video, and the LTX Studio platform with its credit system carry a price tag.
They play different games. Veo 3.1 aims at turnkey cinematic output inside Google's cloud, while LTX-2 bets on open weights, local execution and a per-clip cost of zero once the hardware is paid off. For high volume, sensitive data or fine-tuning, that openness carries real weight.
Yes, through PyTorch's Metal (MPS) backend on Apple Silicon. A MacBook Pro M4 Max with plenty of unified memory handles 4K at roughly half the speed of an RTX 4090, while smaller chips stick to 1080p. LTX Desktop also ships a macOS version, with an optional switch to the API.
Yes, the official repository ships a LoRA trainer so you can adapt the model to a house style, a recurring character or a specific product. The full training code is published, something very few video models provide, closed or open.
Verdict: Zero credit meters and a fully inspectable pipeline: that pitch lands with independent studios churning out clips, product teams wiring video into their apps, and tinkerers with a capable GPU under the desk.
Deevid AI 
Kling 3.0 
MiniMax H3 
