LTX-2.5 generates 10-second video in 6.8 seconds locally, 58x faster than Kling 3.0 Pro

LTX dropped LTX-2.5 today, an open-weights world model purpose-built to run locally on NVIDIA RTX GPUs, and the speed numbers are genuinely disruptive. Published benchmarks show the model generating a 10-second clip in 6.8 seconds on two NVIDIA GB200s. That’s faster than the video’s own runtime and roughly 7.6 times quicker than the fastest closed alternative, Omni Flash, which clocks in at 52 seconds. At the other end of the chart, Kling 3.0 Pro takes 398 seconds. The gap isn’t incremental; it’s the difference between staring at a progress bar and waking up to a batch of finished renders.

The architectural overhaul goes beyond raw speed. LTX rebuilt the pipeline with a new diffusion video decoder that cuts artifacts in high-motion shots and introduced native multishot generation to keep a character’s look consistent across an entire sequence. A custom Gemma 4 language backbone sharpens prompt comprehension, and a technique LTX calls Diffusion Fidelity Rendering builds motion in a compressed latent space before generating high-fidelity keyframes. The result addresses the glitching that made earlier open video models a non-starter for professional campaigns. It all runs inside ComfyUI, meaning a creator can lock a branded character with a quick LoRA fine-tune and never send proprietary IP to the cloud.

For ad teams and short-form creators, the economics flip entirely. Ad fatigue typically hits within 7 to 10 days, but the bottleneck was never coming up with ideas—it was the cost and turnaround time of producing enough assets to test. Local generation erases those per-clip fees and cloud bills. One person on a single RTX workstation can spin up a dozen variations on a brief, localize for five markets, and refresh creative before fatigue sets in. Solo creators can now match the output volume of a full studio.

LTX-2.5 arrives as part of NVIDIA’s month-long local AI push, debuting alongside the open Nemotron 3.5 Lightning agent model. The signal from both releases is clear: open models accelerated on local hardware are becoming default production infrastructure, not just experimental toys. LTX claims the model family has passed 33 million downloads, and this release adds a physical AI checkpoint aimed at robotics simulation. Whether the real-world consistency holds up outside benchmark conditions remains the open question, but the speed alone changes what a single creator can attempt in an afternoon.