Nvidia releases Nemotron 3.5 Lightning, an open 30B-parameter MoE model that it says delivers up to 4x faster output speeds, and an agentic AI model router (Kyt Dotson/SiliconANGLE)

LTX-2.5, the latest open-weights video and world model from LTX, has been launched in partnership with ComfyUI. This new model is now available through ComfyUI, the LTX API, and the Hugging Face repository. LTX claims that LTX-2.5 generates a 10-second, 720p image-to-video clip in just 6.8 seconds, which is faster than real time—but only when run on NVIDIA's top-end GB200 chips.

The model offers multi-shot generation, a diffusion decoder for better quality, new conditioning modes, and improved support for autoregressive models. LTX-2.5 also features a tuned Gemma 4 language backbone and a pretrained checkpoint for physical AI and robotics. The company reports that LTX-2.5 has a 67% win rate in blind, side-by-side human preference tests against other models, though these results are preliminary.

LTX emphasizes the model's flexibility, allowing customers to fine-tune it on their own data and IP. The release is seen as a move against the API business model, as LTX seeks to make video and world models more accessible to businesses of all sizes.