LTX-2.5: Open AI Video With Sound — the Honest GPU Math

LTX-2.5 landed on Hugging Face this weekend from Lightricks, and it is the most interesting open-weights release in months for one reason that has nothing to do with benchmarks: it generates video with synchronized sound — dialogue, ambience, effects matched to the picture — from a text prompt or a single image, under a license that says plainly: if your business makes under $10 million a year, commercial use is free. That combination — open weights, sound included, production use allowed — is the thing the AI-video world has been waiting for, and it is real and downloadable today. Now the honest half, before you clear disk space: the full pipeline is roughly a 66 GiB download, built around a 22-billion-parameter video transformer plus a 12-billion-parameter text encoder, and Lightricks publishes optimization flags rather than a minimum GPU — their attention advice name-drops Hopper and Blackwell, which are data-center architectures. "Open" has never meant "runs on your laptop," and this post exists to do the arithmetic the launch coverage will skip: what LTX-2.5 actually is, what the parameter counts imply about VRAM, the three real ways to use it this week (one of them costs about a coffee), and who should genuinely bother.

What is LTX-2.5, and why is this one different?

Lightricks — the company behind Facetune, and lately the LTX-Video open-model line — released LTX-2.5 with weights on Hugging Face and code on GitHub. Architecturally it is a 22-billion-parameter diffusion transformer (DiT) for the video itself, steered by a customized Gemma-4 12B text encoder, with a video VAE alongside — one safetensors file per component, roughly 66 GiB all-in at full precision. It takes text, an image, an existing video, or even audio as the starting point, and produces picture and soundtrack together — not a silent clip you dub later.

Two capabilities separate it from the wave of video models you have scrolled past. First, the synchronized audio: sound generated with the frames, matched to what is happening in them. Second — and quieter but arguably bigger — native multishot generation: the model can produce a sequence of shots that hold the same character identity, environment, lighting, voice, and visual style across cuts. Every AI-video tinkerer knows the old failure: shot two stars a slightly different person in a slightly different room. Multishot is aimed at exactly that, and it is the difference between "a five-second clip" and "a scene." Individual shots run up to 121 frames — about five seconds at the default 24 fps (frame counts follow an odd rule: they must land on 1, 9, 17… a byproduct of how the model chunks time) — at a default 1024×1536, with settings reaching 4K (3840×2176).

The license: free for real use under $10M — read that again

Most "open" video models arrive with research-only licenses, output restrictions, or user-count tripwires. The LTX-2.x Community License states its terms in one breath: commercial and production use at no cost for anyone under $10 million in annual revenue; above that line, you pay Lightricks for a license. For a freelancer making product clips, a repair shop wanting a thirty-second promo, a YouTuber, or nearly every small business on Earth, that reads as: yours, free, including for paid work. It is the same shape as the deal that made certain game engines ubiquitous — free until you are big enough that paying is a good problem to have. As always: we summarize, the license file on the repo governs; read it before building a business on it, and note that "2.x" means these terms cover this model line, not necessarily the next.

Can your GPU run LTX-2.5? The arithmetic, honestly

Lightricks does not publish a minimum GPU, so let us do what they let us do: count. These numbers are computed from the published parameter counts and standard precision sizes — they are floors, before activations and the memory cost of holding dozens of frames in flight, which for video models is substantial:

Read the table like this. A 16 GB laptop or GPU: no — not the full model, not the fp8 version, not usefully; this is not a 9B LLM situation, and anyone telling you otherwise this week is guessing louder than we are. A 24 GB card (RTX 4090-class) or a 32 GB card: plausibly — the fp8/int8 transformer roughly fits, the encoder offloads to CPU, the repo ships --offload cpu and --offload disk flags for exactly this, and the distilled + NVFP4 variants exist to push it further down. Expect generation to be slow and settings-sensitive, and expect the first two weeks of community ComfyUI workflows to matter more than anything we can tell you today. Above that — the cards Lightricks’ own docs wink at (Hopper, Blackwell) — it simply runs. We have not run it ourselves; our test machine has no GPU worth the name, and this site does not fake benchmarks. What we can promise is that the arithmetic above is straight from the published parameter counts, and it will not bend for enthusiasm.

The three real ways to use LTX-2.5 this week

1. Own big GPU (24 GB+): ComfyUI or the official pipelines. Lightricks ships official ComfyUI workflows — the sane path for most people — plus a ltx-pipelines Python package (GitHub: Lightricks/LTX-2) and a Diffusers repo (Lightricks/LTX-2.5-Diffusers). Start from the int8 Comfy or fp8 variants, not the bf16 files; add --offload cpu before concluding anything does not fit; and give the first run patience — a 66 GiB-class pipeline’s first launch is mostly disk and sighing.
2. No big GPU: rent one by the hour. The type-not-brand version: GPU-cloud marketplaces rent data-center cards per hour for the price of a coffee or two, you run the same ComfyUI workflow there, generate your clips, download them, shut the machine down. For "I need one good thirty-second product video," renting three hours beats buying a $2,000 card by exactly $1,994. This is what we would do, and it is free-license-compatible — the license cares about your revenue, not your hardware.
3. Neither: wait two weeks, on purpose. Every notable open release follows the same arc: launch (today), community quantizations and trimmed ComfyUI workflows (days), then the honest floor settles — someone will publish exactly what the smallest workable card is, with receipts. If your GPU is borderline, letting other people’s VRAM find the edges first is not laziness; it is engineering. This page gets updated when that floor is established — that is a standing promise on every fast-moving post here.

Getting the download right: what the 66 GiB actually contains

The Hugging Face repo is split into one safetensors file per component — the 22B transformer (full and distilled versions), the customized Gemma-4 text encoder, and the video VAE — plus the lower-precision variants. Downloading everything indiscriminately is how a 66 GiB evening becomes a 120 GiB one, so choose before you click:

1. Decide your precision first. 24–32 GB card: take the fp8 or int8 Comfy files, skip bf16 entirely. Data-center card: bf16 if you want maximum quality, fp8 if you want speed and room.
2. Distilled vs full transformer: the distilled version trades some quality for fewer sampling steps — on borderline hardware, distilled is the difference between "an iteration every few minutes" and "a slideshow of your own patience." Start distilled; graduate later.
3. You need all three components — transformer, text encoder, VAE — in matching precision where offered. A missing VAE is the classic "it loaded but outputs noise" forum post of week one.
4. Check your disk before, not during. Weights plus working space plus generated frames want comfortable headroom above the download itself; an SSD matters more than usual because the offload flags page against it.

The launch-day catches nobody puts in the demo thread

Read before you budget an evening. Shots are short by design — 121 frames is about five seconds; the path to longer content is multishot, which is a different craft (thinking in cuts) rather than a longer slider. The frame-count rule bites tools: counts must land on 1, 9, 17… and early third-party workflows that ignore it will error in confusing ways — when something fails, check the frame number before blaming your GPU. Launch demos are the model’s best day: curated prompts, many takes, the winners published; your first-evening hit rate will be lower, and Jake’s folder of hilarious rejects below is the normal experience, not the exception. And the license is generous but versioned — the 2.x Community License covers this line; treat each future major release as a fresh read, not an assumption. None of these are complaints. They are the difference between arriving informed and arriving disappointed, which is most of what this site is for.

Jake wants a shop promo, and the math does its job

Jake saw the launch demos — a talking clip, with sound, from one prompt — and had the entirely reasonable thought: a fifteen-second "we fix cracked screens same-day" video for the shop’s booking page, free, no agency. Then he asked the question this series has trained him to ask first: "What does it need?" Ethan walked him through the table above and put it in shop terms: "This is not a new phone app; it is a new hydraulic lift. Wonderful machine, free to use commercially — and it needs a garage with a reinforced floor. You have a driveway." Jake’s laptop tops out at 16 GB of ordinary RAM and shares graphics memory with the browser. The honest options were the three above, and he picked renting: one evening, a few dollars of GPU time, the official ComfyUI workflow, and a folder of takes — two usable, several hilarious, one where the phone repaired itself in reverse. The usable ones are on the booking page now. Total cost: less than the sign on his door. The lesson is the series’ oldest one wearing new clothes: the model being free was never the question; the question is always what carries it.

Where it sits: LTX-2.5 vs the video-AI landscape

One caution that belongs in every AI-video post and rarely appears: with sound and consistent characters, output from models like this gets convincing fast — and the scam economy adopts convincing tools first. If your family verification for "it’s really me, send money" is still "the voice sounded right," this week is a fine deadline for agreeing on a code word instead. Open tools did not create that problem, but they do retire the excuse that nobody could fake it.

FAQ — LTX-2.5, answered straight

What is LTX-2.5?

Lightricks’ open-weights AI video model: a 22B diffusion transformer with a Gemma-4-based 12B text encoder that generates video with synchronized audio from text, an image, a video, or audio — including multishot scenes that keep character and style consistent across cuts.

Is LTX-2.5 free for commercial use?

Under the LTX-2.x Community License, commercial and production use costs nothing if your annual revenue is under $10 million; above that, you license from Lightricks. The license file in the repo is the governing text — read it before you build on it.

How much VRAM does LTX-2.5 need?

Lightricks publishes no minimum. From the parameter counts: the 22B transformer is ≈44 GB at bf16 and ≈22 GB at fp8 before activations, so 24–32 GB cards are the plausible floor using fp8/int8 variants plus CPU offload, and data-center cards run it comfortably. A 16 GB machine is out for now.

Can I run LTX-2.5 on a 16 GB laptop?

No — not the full model or the fp8 version, and this is arithmetic rather than pessimism. Rent a cloud GPU by the hour for a few dollars, or wait for community quantizations to establish the real floor.

Does LTX-2.5 generate sound with the video?

Yes — synchronized audio generated together with the frames, which is its headline feature among open models. It also accepts audio as an input to drive a video.

How long can LTX-2.5 videos be?

Individual shots run up to 121 frames — about five seconds at the default 24 fps (frame counts must land on 1, 9, 17, and so on). Longer sequences come from multishot generation, which chains shots while preserving character and style.

What resolution does LTX-2.5 support?

Default 1024×1536 at 24 fps, with dimensions divisible by 32, up to 4K (3840×2176) on hardware that can carry it.

What is multishot generation?

Generating several shots as one job so the character, environment, lighting, voice, and style stay consistent across cuts — the failure mode that made older AI video look like a casting change every five seconds is the specific target.

How do I run LTX-2.5 in ComfyUI?

Lightricks publishes official ComfyUI workflows alongside the release — start from those and the int8 Comfy variants rather than the bf16 files, and expect community-trimmed versions within days. The GitHub repo (Lightricks/LTX-2) and the Diffusers package are the code paths.

What are the fp8, int8, and NVFP4 versions?

Lower-precision copies of the weights that roughly halve (or better) memory needs at some quality cost — the same idea as GGUF quantizations for language models. They plus the --offload cpu/disk flags are how smaller GPUs get a chance.

Is LTX-2.5 better than Sora or Veo?

Different question than it sounds: those are closed services you rent access to; LTX-2.5 is weights you possess. On pure output quality the closed frontier still leads; on control, privacy, cost at volume, and the right to run it yourself, open weights win by definition. Pick by which of those you are optimizing.

Can LTX-2.5 animate a photo?

Yes — image-to-video is a core mode: one still becomes a moving, sounding shot. Video-to-video (restyling or extending existing footage) is also supported.

What GPU do the developers themselves assume?

Their optimization notes reference FlashAttention builds for Hopper and Blackwell — data-center architectures — which tells you the native habitat. Consumer cards are the community’s frontier, not the vendor’s baseline.

Is renting a cloud GPU for this actually cheap?

For occasional use, yes — hourly marketplace rates for big cards run in coffee money, and a promo clip’s worth of takes fits in an evening. Buying hardware only wins if you will generate constantly.

Did you run LTX-2.5 yourselves for this post?

No, and we will not pretend otherwise — our test machine has no suitable GPU. Every figure here comes from Lightricks’ own model card, repo, and license, and the VRAM discussion is labeled arithmetic, not benchmark. When we or the community establish the true consumer floor, this page updates.

Should I be worried about scams using models like this?

Alert, not worried: convincing fake video with matching voice is now cheap, so retire "it sounded like them" as your family’s money-transfer verification and agree on a code word. Tools are neutral; the habit is the defense.

Where to go next

Revision note. Written August 31, 2026, within about a day of the LTX-2.5 weights appearing on Hugging Face. Every specification — parameters, modes, frame rules, resolutions, the ≈66 GiB download, quantized variants, offload flags, license terms — was read from Lightricks’ own model card, GitHub repo, and license that day; the VRAM figures are arithmetic from parameter counts and are labeled as such, because we have no GPU that could test them and this site does not invent benchmarks. Three things will go stale fastest: community quantizations will establish the real consumer floor, ComfyUI workflows will improve weekly, and the license terms apply to the 2.x line only — tell me through the contact page when any of those move and I will update this. And if you read the requirements table with a sinking feeling about your own hardware: that feeling is the industry’s, not yours. A year ago this capability did not exist at any price you could download. The floor always comes down. It just never comes down in week one.