LTX-2.5 locally: installation on Mac and Windows, requirements, and first AI video

Can LTX-2.5 really be run on your own computer without paying for each generation? Yes – but only on the right hardware and in the right mode. This guide separates the local video model from cloud services, shows the installation of LTX Desktop on Mac and Windows, the first experiment, and the costs and limitations that short social media videos usually do not explain.

Status as of October 4, 2026. LTX Desktop is in beta. Before downloading several gigabytes, check the latest requirements and license in the official sources linked in the text.

What is LTX-2.5?

LTX-2.5 is a family of models for generating moving images and synchronized sound. You can describe a scene with text, animate an image, or use audio input. The model has open weights that can be downloaded and run locally. "Open weights" do not mean a lack of licensing; the terms differ for small and large entities.

The easiest way to get started is with LTX Desktop, the official application with a graphical interface. It locally runs LTX 2.5 Fast. The Pro version in this application uses the API. More advanced users can work through ComfyUI or the Python repository, but that is a different path with higher requirements and more settings.

Quick answer: is your computer suitable?

These are the thresholds for the Desktop application, not a promise of fast performance. Separate documentation for the full open-source model states that for the best results, a minimum of 32 GB VRAM, 32 GB RAM, 100 GB of space, CUDA 12.7, and Python 3.12 are required. The numbers are not contradictory: the lightweight Fast flow in the application and the fuller Python pipelines have different requirements. On weaker but acceptable hardware, generation may take a long time and limit resolution or clip length.

“Free” has three meanings

1. No fee for each local generation: after downloading the weights, the model computes the film on your GPU. You still pay for hardware, electricity, and time.
2. No cloud at all: in the settings, enable Local Text Encoder and download its additional files. If you use a free LTX key for text encoding, the film can be created locally, but the text prompt will go to the service.
3. No licensing restrictions: this is not true. The model is subject to the LTX-2.x license. The official model card describes free commercial use for entities with less than $10 million in annual revenue; above this threshold, a contract is required. Check the full license for details and definitions.

The API mode for films, the Pro version, and some editing features are paid. When you enable API features, input materials may be sent to the provider. In LTX Desktop, generating and editing single images is a separate function of the application; in API mode, it may use the fal.ai service and the Z Image Turbo model. Do not assume that “LTX-2.5 does images locally” means the same model and the same path as local video. Check the local/API switch for specific operations.

Installation on Windows – step by step

1. Check the card in Task Manager → Performance → GPU. Note the NVIDIA model and dedicated VRAM, not shared system memory. Also check free disk space.
2. Go to official LTX Desktop releases and download the latest Windows installer. Do not download “ready packages” from random videos or drives.
3. Install and run the application. The first launch may download the environment and large model files. Choose a disk with sufficient free space.
4. If the application asks for access to LTX-2.5 weights, log in to Hugging Face and accept the license on the model page.
5. In the settings, check Local and the model LTX 2.5 Fast. If you see “API-only,” return to the hardware requirements. Do not accidentally start paid generation.
6. Select text encoding. A free LTX key can relieve memory, but it sends the prompt to the cloud; Local Text Encoder requires additional downloading and memory but allows for a completely local process.

Installation on Mac – step by step

1. In the Apple menu → About This Mac, check if the processor is Apple Silicon and if the system is macOS 13 or newer. Intel Mac will not run local Desktop mode.
2. Check free memory in Activity Monitor. The application assesses it at startup, not just the size of installed memory. Close heavy programs, especially browsers with many tabs.
3. Download the macOS installer from the official releases page, install and run it. At the first startup, choose a location for the models on a disk with plenty of free space.
4. Log in to Hugging Face and accept the terms of the weights if the application prompts you. In Settings → Models, ensure that LTX 2.5 Fast and local mode are selected.
5. If a Mac with the appropriate memory starts in API mode, close the application, free up RAM, and reopen it. The free memory check occurs only at the start of the process.

Application files and models are usually located in ~/Library/Application Support/LTXDesktop/ on Mac and %LOCALAPPDATA%\LTXDesktop\ on Windows, unless you choose a different model location. Before deleting them, check if they contain projects you want to keep.

Your first video: a simple exercise

Select Text to Video, the Fast model, and local mode. Start with one character, one action, and a stationary camera. Choose a shorter length and the lowest available resolution for testing; only when the movement looks good, increase the settings. A long prompt with many twists can be computationally expensive and hard to evaluate.

Save the result and check three things: whether the crane has the same number of wings throughout, whether the movement is gradual, and whether the sound matches the scene. If the model loses the paper, limit the action to one small movement.

Animate your own image

In Image to Video mode, attach your own photograph or illustration as the starting frame. This is a good way to maintain composition. Describe the movement of what is already visible, rather than adding several new objects. For a photo of a wet street: “the camera is stationary, one tram passes by, reflections ripple in the puddles.” Compare the first frame of the result with the source; if the buildings have already changed shape at the start, simplify the prompt or choose a stronger image anchoring if the interface offers it.

What can LTX-2.5 Fast do locally, and what can't it?

If you have a powerful computer and want more control

The official LTX-2 repository provides Python pipelines, and ComfyUI documentation describes working in nodes. These are options for those who want to set models, adapters, length, input frames, and rendering in detail. The model files for a sample Python setup total about 66 GiB. Full or high-quality flows usually require significantly more memory than the simple Fast in Desktop. Do not start down this path if your goal is just the first short clip.

Common issues that arise

Is it worth it?

Yes, if you have a supported computer and want to experiment with AI video without paying for each local attempt. The biggest advantage is the ability to work on your own weights and control the flow. The biggest costs are memory, disk, and rendering time. If you have a computer without the appropriate card or a 16-gigabyte Mac burdened with other applications, the phrase “free on any laptop” will be misleading.

For further work, our guide on framing and focal lengths and character consistency card will be helpful in evaluating the result, whether you generate locally or in the cloud.