Lightricks has released LTX-2.5, an open-weights world model built to sit under video production, real-time avatars and — increasingly — robotics simulation. The 22-billion-parameter diffusion transformer generates synchronised video and audio in a single pass, and ships alongside a pretrained physical-AI checkpoint that developers can fine-tune on their own data.
Why Robotics Teams Care
Unlike a language model that predicts text, a world model has to represent how an environment changes over time. That makes LTX-2.5 useful for simulation, robot policy pretraining and long-horizon prediction. Markov Robotics is one of the launch partners: co-founder and CEO Atharva Gundawar said his team has not found another open model that models physical behaviour the way LTX-2.5 does and, crucially, one they can run on their own hardware. Real-time infrastructure company Reactor and film studio Asteria are also using the model in production.
What Ships In 2.5
The release rebuilds much of the generation pipeline. LTX highlighted a new diffusion video decoder that reduces artifacts on high-motion sequences, native multishot support that maintains characters and voices across cuts, a Gemma 4-based prompt backbone, native 4K HDR output and a distilled variant that trades minimal quality for lower compute. Weights land on Hugging Face, with native support in ComfyUI and access through the LTX API for managed deployments.
Nvidia-Accelerated, Locally Deployable
Lightricks worked with Nvidia to cut LTX-2.5's memory footprint on RTX GPUs and DGX Spark, so the model can be run across data-centre GPUs, workstations and even Macs. That local-deployment story matters for enterprises hesitant to send proprietary robotics, industrial or media data to a hosted API. LTX-2.5 is free for organisations under $10 million in annual recurring revenue. See related physical-AI moves in LG and Nvidia's Isaac GR00T pact and OpenAI's ultra-fast GPT-5.6 Sol tier on Cerebras.
Reporting based on coverage from The AI Insider, VentureBeat and Lightricks' LTX newsroom.