LTX, a company spun off from Lightricks, officially released the open weights video and world model LTX-2.5 today. The model is now available on Hugging Face and has formed a strategic partnership with ComfyUI for native integration, while also offering managed API services.

On the business model side, LTX adopts a tiered licensing strategy. Organizations with an annual recurring revenue (ARR) of less than $10 million can use it for free, while large enterprises need to negotiate for authorization. The series of models has already exceeded 33 million downloads so far.

QQ20260812-150316.jpg

LTX-2.5 features a deep reconfiguration of the generation pipeline, with core upgrades including: the introduction of a new diffusion video decoder to enhance details and image quality, support for native multi-camera single-pass rendering, the integration of a custom Gemma4 language backbone network, and specialized pre-trained checkpoints tailored for physical AI and robotics technologies. Thanks to NVIDIA optimization, its efficient distillation model can run cost-effectively on hardware such as RTX GPUs and Macs. Official data shows that the model's generation cost is approximately one-eighth of similar models, and the rendering speed is seven times faster.

Performance-wise, in self-hosted tests using two top-tier NVIDIA GB200 chips, LTX-2.5 generated a 10-second 720p video in just 6.8 seconds (faster than real-time playback); the quality blind test win rate reached 67%. LTX CEO Zeev Farbman stated that betting on open-source weights aims to attract developers through openness, and by taking the path of "local verification first, then commercial licensing," it precisely enters diverse enterprise-level application scenarios such as virtual production, real-time robotics, and edge computing.