LTX-2.5 is an open-weight model from Lightricks that generates video and synchronized audio together, and the headline change over LTX-2 is native multishot generation: one prompt can now produce a sequence of shots instead of a single continuous take. The weights are on Hugging Face as Lightricks/LTX-2.5, where the page was trending with more than 8,500 likes on October 20, 2026. This post walks through what the model card actually says, what the license allows, and how to get it running.
The model card publishes no benchmark numbers, so this is a specs-and-setup story rather than a leaderboard story. Where a claim comes from somewhere other than the card, we say so.
TL;DR: what you need to know about LTX-2.5
| Question | Answer (from the model card unless noted) |
|---|---|
| What is it? | A 22B-parameter diffusion transformer that generates video and audio together from text, image or video inputs |
| New vs LTX-2? | Native multishot generation, diffusion video decoder, Gemma 4 12B text encoder, prompt enhancer, optional duration predictor, improved distilled model |
| Is it open? | Open weights, but gated on Hugging Face and released under the LTX-2.x Community License |
| Free for commercial use? | Yes under $10M annual revenue; paid agreement above that |
| How do I run it? | ltx-pipelines in Python, official ComfyUI templates, or a Diffusers pack (install Diffusers from main) |
| Minimum setup? | Python 3.12+, CUDA 12.7+, PyTorch about 2.7; low-VRAM mode uses fp8-cast with CPU offload |
| Benchmarks? | None published on the card |
What is LTX-2.5?
LTX-2.5 continues the LTX line from Lightricks, the company behind the LTX Studio product and the earlier open LTX video models. The card frames it as a model built for local execution and fine-tuning, with the pitch of "full control and customization", meaning you can self-host it on your own infrastructure rather than calling a closed API.
Under the hood the card describes a 22B-parameter DiT (diffusion transformer) shipped as split safetensors files, one per component. The pieces are:
- the DiT itself, in a distilled variant and a full "dev" variant;
- a Gemma 4 12B text encoder with projections, custom to this release;
- a video VAE (a higher-quality DiffVAE or a faster convolutional VAE) and an audio VAE with a vocoder;
- quantized options in bf16, Comfy int8 and NVFP4, some of them ComfyUI-only.
The distilled variant uses a fixed 8-step schedule with CFG set to 1, which is where the speed comes from. The dev variant is the one you train and fine-tune.
What changed from LTX-2?
The card lists six differences from LTX-2. In plain terms:
- Native multishot generation. LTX-2 produced single continuous shots. LTX-2.5 can generate several shots in one go, which matters for anything that looks like a scene rather than a clip.
- A diffusion video decoder. It replaces the VAE reconstruction stage, which is the step that turns latents back into pixels.
- A custom Gemma 4 12B text encoder. Better prompt understanding is the intended gain, though the card gives no measurement. If you follow the Gemma family, our write-up of Gemma 4 26B-A4B on Mac with MLX covers the same model family from the inference side.
- A prompt enhancer. It rewrites a short prompt into the longer, more descriptive form these models respond to.
- An optional duration predictor. If you omit the frame count, the model can set the clip length from the prompt.
- An improved distilled model. The card lists it as improved but gives no measurement.
Lightricks also notes that LoRAs trained on LTX-2.3 mostly work on 2.5, though some need validation. That is useful if you already built style or character adapters.
Resolution, length and audio
The practical constraints are specific. Width and height must be divisible by 32, with examples at 1920x1088 and 960x544 (the first stage in the Diffusers example) and 3840x2176 for 4K via a spatial upscaling pass. Frame count must satisfy frames % 8 == 1, so 121 frames is a valid choice, and the examples run at 24 fps, roughly five seconds per 121 frames.
Audio is generated jointly with video through the audio VAE and vocoder, so sound arrives in the same pass rather than from a separate model, and the card says voice stays consistent across multishot cuts. That joint approach is the same direction taken by closed systems; for a sense of where audio-and-video generation sits in the wider market, see our coverage of ByteDance's Seedance world-model push and fal's Flux 3 Max video price cut.
Open weights being released toward developers, illustrating the LTX-2.5 open-weight video model download
How to run LTX-2.5
The model card offers four routes.
Python with ltx-pipelines. Clone the LTX-2 repository on GitHub, Lightricks' official inference and LoRA trainer package, run uv sync, download weights with hf download, then launch python -m ltx_pipelines.distilled (or the dfr_pipeline module for the full model). Requirements are Python 3.12 or newer, CUDA 12.7 or newer and PyTorch near 2.7.
ComfyUI. Lightricks provides official workflow templates, and the int8 and some NVFP4 checkpoints are ComfyUI-specific. For most hobbyists and creative teams this is the easiest path.
Diffusers. Use the Lightricks/LTX-2.5-Diffusers pack. Support is not in a released Diffusers version yet, so you need to install from the main branch.
Hosted playground. An API Playground exists for trying prompts without a GPU.
For tight VRAM, the card suggests --quantization fp8-cast --offload cpu. The NVFP4 checkpoint with ltx-pipelines needs Blackwell GPUs and the ltx-kernels package, so owners of older cards should stick with fp8 or int8. If you are choosing hardware for local generation in general, our look at the DGX Spark price and memory change is a useful reference point for what a desktop-class Blackwell box costs.
License: open weights, with a revenue line
The weights sit behind a Hugging Face gate: you have to agree to share your account information to download them. The terms are the LTX-2.x Community License:
- commercial and production use is free for entities under $10 million in annual revenue;
- above that threshold you need a paid commercial agreement with Lightricks;
- transferring fine-tunes may also require a paid license.
This is "open weights", not open source in the OSI sense. A solo creator or small studio can ship work made with it. A larger company should read the full license before building a product on top, particularly the clause on fine-tunes, because it affects whether you can hand a trained adapter to a client.
How does it compare with other video models?
With no numbers on the card we cannot rank it, and we will not invent any. What the card does give is a design position: LTX-2.5 is the model you can run yourself, with joint audio, trainable weights and a distilled 8-step mode. Closed video services typically win on peak quality and polish; an open 22B model wins on control, cost at volume, privacy of source footage and the ability to fine-tune on your own style.
If you are weighing the two approaches, a useful reference for how fast video tooling is moving is our piece on how viral Opus 5.5 videos are made with code rather than diffusion, which shows a very different route to moving images, and Google Research's multi-agent long-form video work, which tackles the same consistency problem multishot generation is aimed at.
Who should care about an open video model with audio?
Three groups get the most from a release like this. Independent filmmakers and short-form creators gain a model they can run offline, which matters when source footage is unreleased or client-owned. Studios and agencies with custom looks gain a trainable dev checkpoint, so a house style can be baked into a LoRA instead of being re-prompted on every shot. Researchers and tool builders gain a 22B reference point with documented components, which is rarer than it sounds in a field dominated by closed services.
One timing caution: the model card does not state a release date, and the page shows about 2 million downloads in the last month, so LTX-2.5 may have been available for a while before it started trending. The card also calls its applicability to robotics and physical AI "developing". Check the card, not secondary summaries, for file names and license terms.
Limitations the card states
Lightricks is direct about the weak spots. The model is not a source of factual information, may amplify societal biases, depends heavily on prompt style, may not match prompts perfectly and can produce inappropriate content. Some older LoRAs need re-validation. The Diffusers integration is not yet in a stable release. Multishot is new, and the card does not say how many shots or how long a sequence it handles reliably, so test your own scenes before committing a pipeline to it.
What to do this week
- Request access on the model page and pull the distilled checkpoint first; it is the cheapest way to judge quality.
- Try the ComfyUI template with a short, concrete prompt, then the prompt enhancer with the same idea to see how much it adds.
- Test multishot with a two-shot scene (an establishing shot and a close-up) and check subject consistency across the cut.
- If you plan commercial use, read the license and note the $10M revenue line and the fine-tune transfer clause.
- Wait for independent evaluations before treating it as a drop-in replacement for a paid video API.
Specs, license terms and install steps are as listed on the model card on October 20, 2026 and may change.
Related reading
- Seedance and the world-model race
- Flux 3 Max video model price cut
- How viral Opus 5.5 videos are made with code
- Google Research long-form multi-agent video generation
- Gemma 4 26B-A4B on Mac with MLX
- NVIDIA DGX Spark price and memory change
- Gemini agentic video understanding in AI Studio
Official sources: LTX-2.5 model card on Hugging Face and LTX-2 GitHub repository.