LTX has launched LTX-2.5, the latest version of its open-weights world model, introducing new capabilities for video generation, real-time applications and physical AI.
The company says the new model delivers improvements in visual quality, prompt understanding, generation speed and efficiency, while allowing developers and enterprises to run and customize the model on their own hardware.
LTX-2.5 is available as open weights through Hugging Face, with native support in ComfyUI and access through the LTX API for managed deployments.
LTX says its world models have now recorded more than 33 million downloads and are being used across applications including film production, robotics and real-time rendering.
The company is positioning LTX-2.5 as a foundation model not only for generating video, but also for systems that need to model how environments change over time – an increasingly important capability in robotics and physical AI.
Zeev Farbman, co-founder and CEO of LTX, says: “World models face challenges that LLMs never had to solve, like holding motion, space, and sound consistent across time, which is why efficiency and control matter so much.
“By keeping LTX open, we let teams own their hardware, their IP, and their model. LTX-2.5 is the most capable model yet, and partnerships like the one with ComfyUI, built on a shared belief in an open future, are how it reaches the teams that need both industry-leading quality and full control.”
New model architecture
LTX says it has rebuilt much of the generation pipeline for LTX-2.5 rather than adding new features to the previous architecture.
Among the major additions is native multishot generation, which enables the model to generate a complete sequence while maintaining consistency between characters, scenes and voices across cuts.
Prompt understanding has also been upgraded through a custom Gemma 4 language backbone and a dedicated prompt enhancer designed to interpret more complicated prompts involving multiple subjects.
For visual output, LTX has introduced a new diffusion video decoder intended to reduce artifacts during high-motion sequences while retaining the model’s high compression ratio.
The company has also developed what it calls Diffusion Fidelity Rendering. The approach constructs motion and structure in an 8x temporally compressed latent space while generating higher-fidelity keyframes to preserve important visual details.
The number of keyframes can change according to the complexity of the scene and available computing resources, allowing the system to concentrate processing on parts of a sequence where greater detail is required.
LTX says improvements to its distilled model also enable near-full-model quality with lower computing requirements and faster inference.
Physical AI and robotics
One of the areas LTX is targeting with the new release is physical AI, where world models can potentially be used to help robots learn how environments behave and predict how actions affect their surroundings.
LTX-2.5 includes a pretrained checkpoint specifically tuned for physical AI and robotics. Developers can use the checkpoint as a foundation and fine-tune the model using their own domain-specific data.
Markov Robotics is among the companies working with LTX on robotics applications.
Atharva Gundawar, co-founder and CEO of Markov Robotics, says: “Training robots means teaching them how the physical world actually behaves, not just what it looks like.
“LTX-2.5 is the open model that gets closest to that for us, and being able to run and fine-tune it on our own hardware is what makes it usable for real robotics work. We have not found another open model that does this for us the way this one does.”
World models are attracting growing interest within robotics because they can provide systems with models of how objects, environments and other agents are likely to behave over time.
Rather than predicting the next word, as a large language model does, a world model attempts to predict the next state or “moment” within an environment.
This could make the technology useful for simulation, robot training and planning, as well as conventional applications such as filmmaking, advertising and gaming.
Optimized for local hardware
LTX has also worked with Nvidia to optimize LTX-2.5 for local inference on Nvidia RTX GPUs and DGX Spark systems.
The companies say the optimization reduces memory requirements, making it possible for more developers and businesses to run the model locally rather than relying entirely on cloud infrastructure.
Gerardo Delgado Cabrera, senior director of product for local AI at Nvidia, says: “Local models enable creators and developers to freely explore their ideas thanks to their local GPUs.
“Through close work with LTX to optimize LTX-2.5 and enable it in popular frameworks like ComfyUI, users with Nvidia RTX GPUs and DGX Spark can get the best performance with significantly less VRAM consumption.”
Local deployment also allows organizations to keep proprietary data and intellectual property on their own infrastructure and customize the model for specific applications.
LTX says LTX-2.5 can run on hardware ranging from data center GPUs to Macs.
ComfyUI integration
LTX has partnered with ComfyUI to make the new model natively available within the node-based development environment from launch.
The integration is intended to allow developers to move from experimentation to production workflows without having to transfer their work to a separate proprietary platform.
Yoland Yan, co-founder and CEO of ComfyUI, says: “ComfyUI and LTX share a simple belief, that open models build better tools, faster.
“With LTX-2.5 running natively in ComfyUI, a developer can go from an idea to a working world-model pipeline in an afternoon, and an enterprise can take that same pipeline straight into production on its own hardware. Open is what lets the community move fast and lets businesses build on what it proves.”
The company is also working with Reactor, a development platform for real-time generative video. Reactor is running LTX-2.5 on its low-latency inference infrastructure for applications including interactive avatars, live generated environments and real-time robotics workloads.
Alberto Taiuti, co-founder and CEO of Reactor, says: “The future of media is real-time, and it only works if the underlying model is fast, controllable, and yours to run. Partnering with LTX means we can tune the model to our own infrastructure and bring live, responsive avatars to users worldwide.”
LTX says Asteria is also using its technology for film and video production.
The company expects world models ultimately to provide an infrastructure layer spanning media, simulation, interactive applications and physical machines.
LTX-2.5 is available now through Hugging Face and ComfyUI, as well as through the LTX API. LTX says the model is free to use for organizations with less than $10 million in annual recurring revenue.
