LTX-2.5 makes 10-second AI video in under 7 seconds

LTX-2.5 is an open-weights video and world model designed for blistering inference speeds and native multi-shot storytelling. When self-hosted on two Nvidia GB200 superchips, it can generate a 10-second 720p video clip in just 6.8 seconds, outpacing real-time playback.

What it can do

Instead of stitching individually generated clips together, LTX-2.5 renders entire sequences in a single pass. This native multishot approach holds character identity, scene lighting, and voice consistent across cuts without manual compositing.

The generation pipeline has been rebuilt from scratch. A new diffusion video decoder reduces visual artifacts during rapid camera movement while reconstructing sharp facial features and text. For prompt accuracy, the model uses a custom Gemma 4 language backbone paired with a dedicated prompt enhancer to handle multi-subject scenes. It also ships with a pretrained checkpoint tuned specifically for physical AI and robotics tasks.

Hardware specs and where to run it

You do not need enterprise superchips to get started. Through an optimization effort with Nvidia, LTX released a distilled model with reduced memory footprints that runs locally on Nvidia RTX GPUs and scales down to Apple Silicon Macs.

That headline speed of 6.8 seconds for a 10-second clip requires two steady-state GB200 superchips. Through LTX's managed API, rendering a 1080p clip takes around 23.7 seconds. The API charges $0.09 per second on its Fast tier, bringing a 10-second 720p clip with synchronized audio to $0.90.

The weights are available now on Hugging Face, alongside day-one native integration in ComfyUI. Organizations pulling under $10 million in annual recurring revenue can use it commercially for free, while larger enterprises must negotiate a commercial license.

Why it matters

If you build generative workflows in ComfyUI or need video generation without ongoing per-frame cloud bills, LTX-2.5 is an immediate download. Solving cut-to-cut consistency within an open model removes one of the biggest friction points in AI video production.

We think LTX-2.5 proves open-weights video can compete directly with proprietary platforms on throughput. While claims of cutting costs to one-eighth do not hold against lightweight tiers like Veo 3.1 Lite, rendering high-frame-rate video faster than real time sets a serious benchmark for local production pipelines.