LTX-2.5 Open Weights: Near-Instant AI Video You Can Run Yourself
Lightricks released LTX-2.5 as open weights Aug 11: 22B params, joint audio-video, a 10s clip in seconds. What self-hosting takes, and when hosted LTX wins.
On August 11, 2026, Lightricks put LTX-2.5 on Hugging Face and walked away from the biggest lever a video-model company usually keeps closed: the weights. The model is a 22-billion-parameter dual-stream diffusion transformer that generates video and audio in one forward pass — no separate TTS pass, no post-hoc sync — and on two Nvidia GB200 superchips it turns a still image into a 10-second clip in about 6.8 seconds. (VentureBeat) That's faster than the clip plays back.
This post is about what actually shipped, what "open weights" obligates you to and doesn't, the hardware bill the headline number skips, and where Versely's hosted LTX 2.3 family still wins even with the open version sitting right there.
What Lightricks actually shipped
LTX-2.5 is a follow-on to the LTX-2.3 line, not a new brand — Versely already runs 2.3 in its hosted catalog, more on that below. What changed in 2.5 is architectural. Per ComfyUI Wiki's release writeup, the model adds native multishot: "the model can generate connected scenes — wide, medium, and close-up shots — in a single generation" while holding a character or product consistent across the cuts, instead of you stitching separate generations together by hand. Auto duration predicts how long a described action actually takes and lands the clip there, rather than forcing every prompt into a fixed 5- or 10-second box. And the output target moved to 4K with HDR pipelines built into the official workflows, not bolted on afterward via upscaling.
Lightricks shipped day-one ComfyUI templates for both text-to-video and image-to-video, plus quantized int8 and NVFP4 checkpoints on Hugging Face alongside the full-precision weights — the kind of release hygiene that decides whether a model gets used by Monday morning or sits ignored for a month while the community reverse-engineers a working pipeline.
What "open weights" actually buys you
Open weights isn't the same claim as open source, and LTX-2.5 is a clean example of the difference. You can download the checkpoint, inspect it, fine-tune it and run it in your own pipeline — that part is unambiguous. You don't get the training code or the training data, so you can't reproduce the model from scratch or fully audit what it learned from. It's also a multimodal model in the literal sense — one system handling video and audio jointly instead of two tools passing files to each other — and the weights you can download are the trained result, not the recipe.
The commercial terms sit in the LTX-2.x Community License: free commercial use for any company under $10M in trailing annual revenue, a paid license required above that line. (VentureBeat) That threshold is generous enough to cover essentially every solo creator, agency and small studio reading this — the ceiling is aimed at the handful of companies large enough to build a competing hosted product on top of Lightricks' own work.
The hardware bill the headline skips
The 6.8-second number came from two Nvidia GB200 superchips — a rack-scale inference setup that costs more than most studios spend on cloud compute in a year, not a benchmark you're reproducing on a rented GPU over a weekend. The number that actually matters for a solo operator is the quantized footprint: FP8 quantization brings LTX-2.5 down to roughly 32GB of VRAM. (VentureBeat)
That rules out a single consumer 24GB card at FP8 precision — you're either renting a 40GB+ data-center GPU by the hour, running the more aggressively quantized int8 or NVFP4 builds Lightricks shipped alongside it, or accepting slower CPU-offloaded generation. None of that is prohibitive if you already run a GPU box for other AI video work. All of it is a real setup cost — driver versions, ComfyUI custom nodes, VRAM headroom for the text encoder — that "6.8 seconds" doesn't mention.
When self-hosting wins, and when it doesn't
Self-hosting LTX-2.5 is the right call when you need to fine-tune the model on your own footage or brand, when you're running high enough volume that a metered API stops making sense, or when a client contract requires generation to happen on infrastructure you control. It's a real, permanent capability, and the Community License makes it free at any scale a solo operator or small studio will actually hit.
It's the wrong call when the job is generating a finished video today without provisioning a GPU, managing quantization formats, or debugging a ComfyUI graph from a Discord thread. That's the gap Versely's hosted LTX 2.3 family fills — same lineage, already wired into a prompt-driven pipeline, billed per generation instead of per GPU-hour.
Generating on Versely's LTX 2.3 family
Versely runs three LTX 2.3 models today, and between them they cover most of what a self-hosted LTX-2.5 rig would otherwise be doing for you:
- LTX 2.3 Text to Video Fast — prompt straight to video at 1080p, 1440p or 2160p, 6-20 second durations, with optional native audio generation, at 4 credits.
- LTX 2.3 Image to Video Pro — animate a reference image (a product shot, a character turnaround) across the same resolution and duration range, plus end-image transitions for a controlled finish.
- LTX 2.3 Retake Video — point at an existing clip, describe what's wrong with one segment ("change the flower to a red rose"), and regenerate just that section instead of the whole render.
A concrete run: open Versely's AI video generator, select LTX 2.3 Text to Video Fast, and prompt something like "Product on a marble counter, slow push-in, warm morning light, 10 seconds, 1080p, native audio of ambient kitchen sound." If the last two seconds land wrong — a prop moves oddly, the light flickers — don't regenerate the whole clip. Feed it into LTX 2.3 Retake Video with a start time and a duration covering just the bad segment, and it replaces that section in place. That retake loop is the practical equivalent of the iterative regeneration a self-hosted rig gives you, minus the VRAM math. Current rates for all three sit on Versely's pricing page.
FAQ
Is LTX-2.5 the same model Versely runs? No. Versely's hosted catalog currently runs the LTX 2.3 family. LTX-2.5 is Lightricks' newer open-weights release; whether and when it lands in a hosted catalog is a separate question from whether you can self-host it today.
Can I use LTX-2.5 commercially for free? Yes, under the LTX-2.x Community License, for any company under $10M in annual revenue. Above that threshold, Lightricks requires a paid license. (VentureBeat)
What GPU do I actually need? Roughly 32GB of VRAM at FP8 precision for the full model — a single consumer 24GB card won't fit it at that precision. The int8 and NVFP4 quantized checkpoints Lightricks shipped alongside the FP8 build bring the footprint down further, at some quality cost.
Does LTX-2.5 generate audio too? Yes — it's built as a joint video-and-audio model, generating synchronized sound in the same forward pass rather than adding a TTS or SFX layer afterward.
Takeaway
LTX-2.5 is a genuine leap for anyone with the GPU budget and the patience to run their own inference stack — native multishot and joint audio-video generation in a 22B open-weights model is not a small thing to give away for free under a $10M revenue line. But "you can run it yourself" and "you should" are different sentences. If the job today is shipping a finished product video, Versely's LTX 2.3 family gets you there without a VRAM budget — and if the job six months from now is fine-tuning a video model on your own brand's footage, LTX-2.5's weights are already sitting on Hugging Face waiting.