What Is LTX 2.5? Free Open-Source AI Video Generator

LTX 2.5 is an AI video model with publicly available weights that generates picture and sound at the same time. You can run it on your own computer, with no per-clip fee and no watermark. That makes it one of the first models people look at when they search for free AI video generation. But "free" comes with three details: the license, the hardware, and the electricity. I cover all three below.

This is the first of a four-part LTX 2.5 series. In order: what it is (this post), how to install it, requirements and speed, and your first video, prompts and settings.

This post is a compilation; I have not tried the model on my own computer. The information comes from the model's Hugging Face page, LTX's official documentation, and independent write-ups. Sources are at the end.

What is LTX 2.5?

LTX is the video model family from LTX, the company spun out of Lightricks. LTX 2.5 is the latest member of that family. OpenSourceForU's news piece places the model in August 2026.

Feature	Value
Size	22 billion parameters, dual-stream diffusion transformer (DiT)
Output	Simultaneous video and audio
Input	Text, image, video
Resolution	720p up to native 4K
Clip length	6 to 20 seconds
Text encoder	Custom fine-tuned Gemma 4 12B
Distribution	Open weights on Hugging Face, ready-made templates in ComfyUI

What the model card lists as new compared with the previous version:

• Multi-shot generation. It produces several connected shots in one pass; character, setting, lighting, sound, and visual style are kept consistent across shots. Earlier versions produced a single continuous shot.
• New diffusion video decoder. Sharper results on faces, textures, and on-screen text, with fewer artifacts in difficult scenes.
• Diffusion Fidelity Rendering. It distributes computation according to the complexity of the scene.
• Gemma 4 text encoder. It handles long prompts with multiple characters, camera movement, and lighting descriptions better.
• Prompt enhancer. It expands a short sentence into a more detailed, cinematic description.
• Duration predictor. An optional node estimates the clip length from the prompt and sets the frame count itself.
• Better distilled model. A smaller, faster checkpoint that keeps more of the full model's visual quality.

What does "free" mean?

Using it without paying

When you download the weights from Hugging Face and run them on your own hardware, you pay no per-clip fee. One independent write-up says "no per-clip fee, no watermark." Generating through the API is paid.

License: open weights, not open source

The model ships under the LTX-2.x Community License. This license does not meet the Open Source Initiative's definition, so "open-weight" is more accurate than "open source." What you need to know:

• Organizations with annual revenue under $10 million can use the model commercially, run it on their own servers, and fine-tune it.
• Organizations with revenue of $10 million or more must sign a paid commercial license agreement. Revenue is calculated across the entire legal entity, including affiliates.
• There is an acceptable use policy: areas such as child sexual abuse, non-consensual sexual content, and explicit pornography are prohibited.
• The files on Hugging Face are gated: you have to accept the license and sign in.

What is binding is the LICENSE file in the repository. If you plan to use it for a commercial job, read it yourself.

Hardware: the real cost

The weights are free, but the graphics card you run them on is not. The official ComfyUI page asks for a CUDA graphics card with 32 GB or more of VRAM and 100 GB of free disk space. The community reports it also runs with less VRAM, but slowly. I go into the details in the requirements post. If you don't have a suitable card, renting a GPU by the hour or using the LTX API are options too.

Who is it for?

• People who want to generate on their own computer and keep control of their privacy. Your images and prompts never leave your machine.
• People who already use ComfyUI or are willing to learn it. The node-based interface is the easiest route.
• People who want to fine-tune on their own data. The "dev" checkpoint is fully trainable; most LoRAs trained on LTX-2.3 are reported to work on 2.5 without changes.
• People who make short clips with sound. Audio and picture come out of a single model.

If you don't want to install anything, cloud-based tools are a better fit. I described the general workflow in the complete guide to making video with AI, and a ready-made cloud tool in the Higgsfield post.

Limitations

The model card's own list: the model is not designed to provide factual information, it can amplify societal biases, it is very sensitive to prompt style, it may not match the prompt exactly, and it can produce inappropriate content. I would add three practical notes: setup requires ComfyUI knowledge, the files are tens of gigabytes, and speed depends tightly on hardware.

Frequently Asked Questions

Is LTX 2.5 really free?

If you download the weights and run them on your own computer, there is no per-clip fee. However, the community license has a revenue cap ($10 million in annual revenue), and the hardware cost is yours.

Is LTX 2.5 open source?

The weights are open, but the license does not meet the OSI definition. The correct term is "open-weight."

Does LTX 2.5 generate audio too?

Yes. Video and audio are generated in the same model, with a separate audio VAE.

How many seconds of video can it generate?

Sources mention clip lengths of 6 to 20 seconds. In practice, the length depends on your graphics card.

Can I use it commercially?

Yes, if your annual revenue is under $10 million. If it is above that, you need a paid license.

Sources

• Lightricks, LTX-2.5 model card (Hugging Face): new features, checkpoints, usage, and limitations.
• LTX, Using ComfyUI with LTX: prerequisites, file list, templates.
• Open Source For You, LTX-2.5 Brings Open-Weight Video Generation: technical specs, release timing.
• The Rundown, LTX-2.5 Review: Features, API Pricing & License: license revenue cap and acceptable use policy.
• Pexo, What Is LTX-2.5?: note on watermarks and per-clip fees (third party).