LTX-2.5LTX-2.5: the open-weights video model, explained
LTX-2.5 is an open-weights AI video model released on 11 August 2026 by LTX, the company spun out of Lightricks. It turns text or an image into video with synced audio, in two tiers: Fast (up to 4K and 20 seconds) and Pro (up to 1080p and 10 seconds). The weights are on Hugging Face and are free for commercial use for organisations under $10M in annual revenue.
| Developer | LTX (spun out of Lightricks) |
|---|---|
| Released | 11 August 2026 |
| Inputs | Text, image, video |
| Audio | Synchronised audio in the output |
| Fast tier | 720p to 4K; up to 20 s at 24/25 fps (10 s at 48/50 fps, 1440p and 4K) |
| Pro tier | 720p or 1080p; 6, 8 or 10 s |
| Speed (vendor claim) | 10 s of 720p image-to-video in 6.8 s on two NVIDIA GB200 GPUs |
| Self-hosting | About 16 GB of GPU memory; a distilled model runs on NVIDIA RTX cards |
| Weights | huggingface.co/Lightricks/LTX-2.5 |
| License | LTX-2.x Community License: free commercial use under $10M annual revenue, paid licence above; not OSI open source |
| API price (vendor) | Fast $0.09/s at 720p up to $0.30/s at 4K; Pro $0.12/s at 720p, $0.17/s at 1080p |
Updated October 27, 2026 · Figures from the vendor and public reports, see sources below
What is new in LTX-2.5
- Native multi-shot generation that keeps characters, scenes and voices consistent across cuts.
- A new diffusion video decoder that reduces artefacts in fast motion.
- A Gemma 4 language backbone with a built-in prompt enhancer.
- Fine-tuning on your own data is allowed under the licence.
Is LTX-2.5 good for image to video?
- Its main strength is speed and cost: it is one of the fastest models that can run on your own hardware, which makes it a strong choice for quick previews and high-volume clips.
- Quality figures published so far come from the vendor's own blind tests (a 67% win rate) and have not been independently verified; public arena leaderboards had not scored it at the time of writing.
- The licence requires you to disclose that output is machine-generated and not remove provenance features.
How to use LTX-2.5
- Run it yourself: download the weights from Hugging Face and use the day-one ComfyUI integration.
- Use an API: LTX's managed API and several model hosts offer Fast and Pro tiers, billed per second of output.
- On Vivimo: free previews run on a fast, low-cost open-weights model, and full-length videos use cinematic models. Upload a photo on the home page to try it.
Try image to video on Vivimo
Turn any photo into a short AI video. 2 free previews a day, no sign-up.
Try it freeFAQ
Is LTX-2.5 open source?
It is open-weights: you can download and run it. The LTX-2.x Community License is not an OSI-approved open-source licence, and companies above $10M annual revenue need a paid licence.
What is the difference between LTX-2.5 Fast and Pro?
Fast is the distilled, speed-focused tier with up to 4K and 20 seconds. Pro aims for higher quality and is limited to 1080p and 10 seconds.
Does LTX-2.5 add a watermark?
No visible watermark, but the licence requires disclosing machine-generated content.
What hardware do I need?
The vendor quotes about 16 GB of GPU memory for self-hosting. A distilled version runs on NVIDIA RTX graphics cards.