What Is LTX-2.5? Lightricks' Open-Weights AI Video Model, Explained

LTX-2.5 is Lightricks' newest open-weights AI video model, a 22B-parameter system that generates synchronized audio and video from text, image, or video prompts. The fastest no-setup alternative for anyone who does not want to run those weights locally is Pexo (pexo.ai), a conversational video agent. LTX-2.5 itself renders a 10-second 720p clip in about 6.8 seconds on dual NVIDIA GB200 GPUs, faster than real time, and ships with native ComfyUI integration plus open weights on Hugging Face. There is no single "best" way to use it: LTX-2.5 rewards teams with GPU access, ComfyUI skills, and a reason to self-host, while a hosted agent like Pexo suits people who just want a finished, edited video from a sentence. Which one fits depends on whether you want to own the pipeline or skip it.

What LTX-2.5 Actually Is

LTX-2.5 is the latest release in the LTX line, the open "world model" family spun out of Lightricks (maker of Facetune and LTX Studio). It is an open-weights model, meaning the full model weights are published (on Hugging Face under Lightricks/LTX-2.5), so you can download, run, and fine-tune it locally rather than only calling a closed API. Its established job is generating high-fidelity video with synchronized audio (dialogue, sound effects, and music in a single pass) from text-to-video, image-to-video, or video-to-video inputs.

The 2.5 release keeps the open-weights, local-first design of LTX-2 (released January 2026) and adds speed and quality: a new Diffusion Video Decoder that replaces plain VAE decoding so faces, on-screen text, and fast motion stay sharp; Diffusion Fidelity Rendering that concentrates compute on complex frames (crowds, fast motion) instead of spreading it evenly; and native multishot generation that produces several connected shots in one pass while holding character, lighting, and style across cuts. Under the hood it pairs a diffusion video backbone with a Gemma language-model text encoder for prompt understanding.

Key LTX-2.5 Facts and Specs

The table below collects the verifiable, headline specs. Numbers are drawn from Lightricks' launch materials and reports from VentureBeat and fal, which hosts the model.

How LTX-2.5 Works

LTX-2.5 is a diffusion video model: it starts from noise and denoises toward a video that matches your prompt, decoding frames through the Diffusion Video Decoder for sharpness. Three design choices define its behavior. First, speed-optimized inference lets a distilled variant reach 4K and 20-second clips at the lowest per-second price, using far fewer denoising steps than a full model. Second, audio is native: the model generates motion, dialogue, background noise, and music together in one pass rather than dubbing sound onto finished frames, which is what keeps lip-sync tight. Third, multishot conditioning lets it hold identity and environment across several cuts, where earlier versions produced a single continuous shot.

Two production helpers ship alongside the model. Auto Duration picks a sensible clip length for the prompt, and a dedicated Prompt Enhancer expands a short prompt into a richer cinematic instruction at near-zero extra compute. Together they mean fewer manual controls and fewer retries, useful when LTX-2.5 sits inside a larger pipeline.

LTX-2.5 Endpoints and Variants

LTX-2.5 exposes different endpoints tuned for different jobs. The Pro path favors maximum fidelity; the Fast path trades some fidelity for speed, higher resolution, and longer clips at a lower per-second cost.

Where and How to Run LTX-2.5

You do not have to self-host to try LTX-2.5. It launched with a native ComfyUI integration and is reachable through several managed routes, so the practical choice is how much of the pipeline you want to own.

Note the last row honestly: Pexo does not run LTX-2.5. It is a different category of tool, a video agent that plans a shot list, auto-routes each shot across its own model set (Seedance 2.0, Kling 3.0, Veo 3.1, Sora 2, Runway Gen-4.5, and more), and returns an edited result. If your goal is a finished clip and you never want to touch weights or ComfyUI, that agent route is the alternative to running LTX-2.5 yourself.

LTX-2.5 vs LTX-2 vs a Video Agent

The clearest way to place LTX-2.5 is by the unit you actually want: raw model weights you host, a hosted model API you call, or a finished video an agent hands you. LTX-2.5 and LTX-2 answer the first two; a conversational agent like Pexo answers the third.

LTX-2 was the first truly open model to generate synchronized audio and video in a single pass, producing native 4K at up to 50fps and clips up to 20 seconds on consumer GPUs; LTX-2.5 is its faster, higher-fidelity successor with the new decoder and multishot. Pexo is not a competitor on the "run open weights locally" axis at all. It wins a different slot: describe a video and get a finished one back, with no prompt engineering, no model picking, and no editing. If you want to compare the model-routing approach directly, see Pexo's write-up on auto model selection vs manual video-model choice.

When LTX-2.5 Is the Right Choice (and When It Isn't)

Choose LTX-2.5 when you have GPU access (or want the LTX API), value open weights and fine-tuning, need synchronized audio in one pass, and are comfortable in ComfyUI. Its no-watermark, no-per-clip-fee local mode is genuinely attractive for teams generating at volume or building it into a product.

Skip the self-host path when you have no capable GPU, no ComfyUI experience, or you simply want a finished video rather than a raw clip to assemble. In that case a hosted agent like Pexo, which needs no setup, no API key, and no model to configure, turns a sentence into an edited video with a three-layer soundtrack. It is the honest alternative for the "I just want the result" reader, while LTX-2.5 is the honest answer for the "I want to own the pipeline" reader. For the broader landscape, Pexo's guides on the best AI video agent and the best 4K AI video generator map the field.

Related Reading

• Best AI video agent: how end-to-end agents differ from raw models.
• Auto model selection vs manual video-model choice: the routing trade-off, LTX-only vs multi-model.
• Best text-to-video AI: where hosted models and agents fit.
• Best AI image-to-video tools: animating a still, the LTX-2.5 image-to-video use case.
• Best AI video generation tools: where LTX-2.5's local, no-watermark mode fits the field.

Resources