LTX 2.5: How to Download and Run This Open-Weight Video Model Yourself
LTX 2.5 is a downloadable 22B open-weight AI video model with synced audio and multi-shot scenes. Here's how self-hosting it actually works.
What is LTX 2.5?
LTX 2.5 is an AI video model from Lightricks that generates video with synchronized audio and supports multi-shot scenes, and unlike most frontier video models, you can download its weights and run it on your own GPU instead of only accessing it through a hosted API. It’s published on Hugging Face as a 22B-parameter diffusion model (Lightricks/LTX-2.5), available in both a full “dev” version and a faster distilled version, with quantized formats included for lower-resource setups.
TL;DR
- Open weights mean LTX 2.5 can be downloaded from Hugging Face and run on infrastructure you control, not just accessed through Lightricks’ website or API.
- The model ships as a 22-billion-parameter transformer, with both a bf16 full-precision version and int8/nvfp4 quantized versions for smaller GPUs.
- A distilled variant trades some quality for speed, generating in far fewer diffusion steps, which makes it the more practical choice for self-hosting.
- LTX 2.5 generates video and audio together in one pass, including dialogue, lip sync, and environmental sound, rather than bolting audio on afterward.
- The release adds native multi-shot generation, letting one generation produce a wide shot, medium shot, and close-up of the same character while keeping face, clothing, lighting, and voice consistent.
- A new diffusion video decoder specifically targets the weak points of AI video: faces, fine textures, on-screen text, and fast motion.
- The download includes supporting components (a Gemma-based text encoder, separate audio and video VAEs, latent upscalers, and a LoRA), not just the core transformer.
Other agents start typing. Remy starts asking.
Scoping, trade-offs, edge cases — the real work. Before a line of code.
How is LTX 2.5 different from closed video models?
Most well-known AI video tools, whether from OpenAI, Google, or Runway, are closed systems. You submit a prompt, their servers run the generation, and you get back a video. You never see the weights, can’t inspect how the model works, and can’t run it anywhere except their platform. If the company changes pricing, rate limits, or availability in your country, your workflow changes with it.
LTX 2.5 breaks from that pattern by being released as open weights. The same model Lightricks offers through its hosted API and managed inference is also published on Hugging Face for anyone to download. That means developers can run it on their own GPUs, integrate it into their own products, fine-tune it for specific use cases, and build custom tooling around it, all without depending on a third party’s uptime or terms of service. The model has already become one of the more downloaded open video models on Hugging Face, with the repository showing over two million downloads and thousands of likes.
What’s actually in the download?
The Hugging Face repository for Lightricks/LTX-2.5 isn’t a single file. It’s a full model package:
- Diffusion transformer weights: both the standard “dev” model and the “distilled” model, each available in bf16 (full precision), int8 (quantized for ComfyUI), and in the distilled case, nvfp4 (a more aggressive quantization for lower VRAM usage).
- Text encoder: a Gemma-based 12B text encoder with projection layers, also available in bf16 and int8.
- VAEs: separate variational autoencoders for audio and video, plus a convolutional variant of the video VAE.
- Latent upscalers: spatial and temporal upscaling models that double resolution and frame count in latent space.
- A LoRA: a distilled 450-step LoRA for further tuning the distilled model.
- A duration head model patch: used to control generation length.
This matters for self-hosting because it tells you what you’re actually running. It’s not just a video generator, it’s a full pipeline: text encoding, diffusion, audio/video decoding, and optional upscaling, all shipped as separate, swappable components. That modularity is typical of how open video models are built for tools like ComfyUI, where each piece can be mixed, replaced, or quantized independently depending on your hardware.
Why does the distilled version matter for running it yourself?
Distillation trains a smaller or faster model to approximate the output of a larger one in far fewer steps. For LTX 2.5, the distilled transformer is built to generate video in significantly fewer diffusion steps than the standard “dev” model, which directly reduces the compute and time needed per generation.
Remy doesn't write the code. It manages the agents who do.
Remy runs the project. The specialists do the work. You work with the PM, not the implementers.
For anyone self-hosting on their own GPU rather than renting a company’s inference cluster, that’s the practical difference between a model that’s usable and one that isn’t. The standard model gives you the highest quality ceiling. The distilled model gives you something you can actually run without a data-center-scale setup, and it comes with quantized nvfp4 weights specifically aimed at squeezing the model onto less VRAM. A creator testing LTX 2.5 locally used the distilled version for exactly this reason: it’s the version built for running outside a managed environment.
How does the audio and video generation work together?
A persistent weakness in AI video has been audio: either there’s no sound at all, or audio gets generated separately and synced afterward, which tends to produce dialogue that doesn’t quite match mouth movement or ambience that feels disconnected from the scene. LTX 2.5 addresses this by generating video and audio as part of the same system, with separate but coordinated VAEs for each. That lets a single prompt describe dialogue, environmental sound, and visual action together, and have the model produce them in sync.
In practice, this shows up most clearly in close-up talking shots, where lip sync, facial consistency, and voice timing are easy to evaluate and hard to fake. It’s also the area where failures are most obvious: a face that shifts mid-sentence, a mouth that doesn’t match the words, or a voice that feels pasted onto the video rather than part of it.
What is native multi-shot generation?
Before this kind of capability, producing a short scene with a wide shot, a medium shot, and a close-up meant generating each shot separately and hoping the character, outfit, lighting, and room stayed the same across all three. In practice, that consistency rarely holds; AI video models tend to generate a “different person” or a reshaped room the moment you ask for a new angle.
LTX 2.5’s native multi-shot generation lets a single generation contain multiple connected shots while preserving character identity, environment, lighting, and voice across the cuts. Instead of treating each shot as an independent generation task, the model treats the sequence as one coherent unit. This is a meaningful shift for anyone trying to build actual scenes rather than isolated clips, since continuity across cuts has been one of the more stubborn problems in generative video.
What does the new video decoder fix?
LTX 2.5 introduces what Lightricks calls a diffusion video decoder, responsible for turning the model’s compressed internal representation back into full video frames. This is the stage where quality problems in AI video usually surface: faces that melt under a closer look, background signage with garbled pseudo-text, or fast-moving objects that blur into mush.
The new decoder is aimed specifically at improving those failure points: faces, textures, text, signage, and motion. These are also the areas worth checking when evaluating any new video model release, since broad, slow-moving shots (landscapes, establishing shots) tend to look good across most models regardless of underlying quality. The harder test is a static close-up with legible text or quick movement in frame.
Is self-hosting LTX 2.5 worth it?
Remy is new. The platform isn't.
Remy is the latest expression of years of platform work. Not a hastily wrapped LLM.
That depends on what you need. If you’re prototyping or generating occasional clips, a hosted API is simpler: no GPU to manage, no weights to download, no dependency chain to maintain. But if you’re building a product around video generation, need to control costs at scale, want to fine-tune the model for a specific style or use case, or simply don’t want a single company able to change pricing or availability under you, self-hosting is the advantage open weights give you that closed APIs don’t.
The tradeoff is infrastructure. Running a 22B-parameter diffusion model, even the distilled and quantized versions, still requires a capable GPU. The availability of int8 and nvfp4 quantized weights lowers that bar somewhat, but this isn’t a model you run on a laptop. For teams with existing GPU infrastructure or access to cloud GPU instances, though, the ability to run the exact same model Lightricks hosts, on hardware you control, under the model’s license, is the core value proposition of the release.
Frequently Asked Questions
What is LTX 2.5?
It’s an open-weight AI video generation model from Lightricks, built around a 22-billion-parameter diffusion transformer, capable of generating video with synchronized audio and multi-shot scene continuity.
Can I actually download LTX 2.5, or is it only available through an API?
You can download it. The full model, including the transformer weights, text encoder, VAEs, and upscalers, is published on Hugging Face (Lightricks/LTX-2.5) under its license, alongside Lightricks’ own hosted API and managed inference options.
What’s the difference between the standard and distilled LTX 2.5 models?
The standard (“dev”) model targets maximum quality. The distilled model is trained to generate in far fewer diffusion steps, making it faster and less compute-intensive, which makes it a better fit for self-hosted or resource-constrained setups.
Do I need a powerful GPU to run LTX 2.5 locally?
Yes, though quantized versions (int8 and nvfp4) reduce the hardware requirement compared to the full bf16 weights. It’s not a model suited to consumer laptops, but it’s designed to run on GPU infrastructure you control rather than requiring a data center.
What makes LTX 2.5’s audio different from other video models?
Instead of generating video and then adding audio separately, LTX 2.5 generates video and audio together through coordinated audio and video VAEs, allowing dialogue, lip sync, and environmental sound to stay coherent with what’s happening on screen.
