Lightricks' LTX-2.5 open video model makes picture and sound together

Lightricks’ LTX-2.5, a 22-billion-parameter open-weights model, generates video and its soundtrack together in one pass from a text prompt or a start image, according to a post on X on November 11 that put the model back in front of developers. The same post says running it locally needs roughly an 80 GB GPU, a figure that does not appear in Lightricks’ own documentation.

The model itself is not brand new. Lightricks’ official GitHub repository says LTX-2.5 was released on August 11, 2026, and that it supersedes both the original LTX-Video (v0.9.x) and LTX-2. What the November attention adds is a plain summary of the trade-off: synchronized picture and sound from a downloadable model, at the cost of serious hardware. This article sets out what is confirmed by Lightricks’ model card and repository, what comes from independent coverage, and what rests on the X post alone.

What LTX-2.5 adds over LTX-2

Lightricks describes LTX-2.5 as the latest LTX model and recommends it for all new projects. Its repository lists four headline changes: native multi-shot generation, synchronized audio and video, stronger prompt adherence and better local performance.

Synchronized sound was already the main pitch of LTX-2, which Lightricks announced on October 23, 2025. At the time, the company said visuals and sound in LTX-2 are generated “in one coherent process, with motion, dialogue, ambience, and music flowing simultaneously”. That announcement also claimed native 4K at up to 50 frames per second, clips with synchronized audio of up to 10 seconds, and “up to 50% lower compute cost than competing models”. Those are Lightricks’ own claims about LTX-2, not independent measurements, and the sources do not restate them specifically for LTX-2.5.

The Rundown, which publishes an independent tool overview, describes the main upgrade in LTX-2.5 as native multi-shot generation: one run can create several connected shots “while attempting to preserve characters, environments, lighting, visual style, and voice across cuts”. It lists three further changes:

• A diffusion video decoder. Lightricks replaced the previous reconstruction stage with a diffusion decoder intended to improve faces, texture, text, motion and pixel-level detail.
• Automatic duration. The model can pick a clip length from the prompt instead of requiring a fixed duration.
• Audio-to-video. Besides generating sound, the model can drive video from an input audio track, such as an existing voice or soundtrack.

The Rundown lists the inputs as text, image, video and audio, which is broader than the text-or-start-image description in the X post.

Inside the downloadable checkpoints

The clearest technical detail comes from Lightricks’ Hugging Face model card for LTX-2.5 Pre-Trained. That checkpoint is the base model produced during pre-training, before the supervised fine-tuning and other post-training used for the main LTX-2.5 release. Lightricks says it is intended for researchers and developers building specialised models, and points general users to the main LTX-2.5 checkpoint instead.

The core of the model is a 22B DiT, short for diffusion transformer: a transformer network that produces images or video by gradually removing noise. In the pre-trained release, that transformer ships as a single 43 GiB file in bf16, a 16-bit number format common for model weights. The same file also contains the video VAE, the audio VAE, a vocoder, a duration head and text-embedding projections. A VAE (variational autoencoder) here compresses and decompresses video or audio so the transformer can work on a smaller representation; a vocoder turns that audio representation back into a waveform.

Text understanding comes from a separate component: a Gemma4 12B text encoder in bf16, which the model card marks as required. In the desk’s reading, packing the video and audio decoders into the same file as the transformer fits the model’s joint audio-video design, but the model card itself only lists the components and does not explain the choice.

The main release is structured differently, according to The Rundown, which says the model card provides both distilled and trainable 22B transformer checkpoints, plus separate text encoder, video decoder, audio, duration and upscaling components. A distilled model is a version trained to reach similar output in fewer steps, which usually makes generation faster. Lightricks’ Hugging Face collection for the family lists base models, quantized models, and accompanying LoRAs and IC-LoRAs. LoRAs are small add-on weight files that adjust a model’s style or behaviour without retraining all of it.

Spec	Value	Source
Release date	August 11, 2026	Lightricks GitHub
Model size	22B diffusion transformer	Lightricks model card
Pre-trained checkpoint	43 GiB, bf16, single bundled file	Lightricks model card
Text encoder	Gemma4 12B (bf16), required	Lightricks model card
Inputs	Text, image, video, audio	The Rundown
Hosted API output	Fast up to 4K; Pro up to 1080p	The Rundown
Local options	Python, ComfyUI, Diffusers	The Rundown
Local GPU memory	Roughly 80 GB (unconfirmed)	X post

Running LTX-2.5 locally: the 80 GB question

The X post calls hardware the “catch”: it says running LTX-2.5 locally needs roughly an 80 GB GPU. Neither Lightricks’ model card nor its GitHub repository, as provided to the desk, states a memory requirement, so the figure should be treated as the poster’s estimate rather than an official spec.

The official numbers point in the same direction without settling it. The pre-trained checkpoint alone is 43 GiB in bf16, and it also needs the 12B-parameter Gemma4 text encoder. Lightricks’ collection includes quantized models, meaning versions stored at lower numeric precision to save memory, but the sources do not give their sizes or how much they lower the hardware bar. Lightricks does say LTX-2.5 brings “better local performance” than its predecessors, without publishing figures in the sources.

For builders, the practical reading is that self-hosting the full model is a job for data-centre-class GPUs, not a typical consumer card, unless quantized variants prove good enough in testing.

How to get LTX-2.5: weights, gating and the hosted API

Lightricks publishes the LTX-2.5 code and weights in its LTX-2 GitHub repository, and its older LTX-Video repository is no longer under active development. The weights are also on Hugging Face, where access is gated: users must log in and agree to share their Hugging Face information, and accept Lightricks’ privacy policy and consent to receive offers and updates, before downloading.

The Rundown lists three local routes: Python pipelines, ComfyUI (a node-based interface popular for image and video workflows) and an early Diffusers integration. Fine-tuning runs through Lightricks’ trainer; the pre-trained model card says to fine-tune it with the LTX-2 Trainer, paired with the Gemma4 text encoder.

For teams that do not want to manage GPUs, Lightricks also runs a hosted API and playground, according to The Rundown. It describes two hosted models: Fast, which supports more resolutions and goes up to 4K, and Pro, which targets higher fidelity but currently tops out at 1080p. The Rundown also notes that the hosted 2.5 API omits several editing endpoints, so not every capability of the open weights is reachable through the hosted service.

The licence: free below $10M revenue, paid above

LTX-2.5 is open-weights, not unrestricted open source. The model card says organisations with under $10M in annual revenue get commercial and production use at no cost under the LTX-2.x Community License. Organisations above that line need a paid Commercial Use Agreement, which Lightricks says comes with full weights, engineering support, LoRAs and flexible deployment options. The model card also warns that transferring fine-tunes may require a paid licence.

The Rundown is direct on the point: the model “should not be described as unrestricted open source,” and revenue counts include entities under common control. It advises teams to review the binding licence before building it into a product.

Fine-tuning drew an early question. On release day, a Hugging Face user asked whether a community fine-tune is allowed without a commercial licence, noting that the 2.5 licence is worded differently from 2.3. A member of the Lightricks organisation replied for the LTX Team:

Yes — community fine-tuning is allowed and welcomed without a paid commercial license, as long as your entity (aggregated with any subsidiaries or affiliates under common control) has annual revenue under $10,000,000.

The reply adds that the resulting weights count as a “Derivative” under the same free terms, but that anyone sharing a fine-tune must pass along the full licence and use restrictions, and that a recipient with $10M or more in revenue needs their own paid licence. That is a forum answer, not the licence text, so the binding licence still governs.

What LTX-2.5 means for builders, and the open questions

The Rundown calls LTX-2.5 “unusually flexible”. It lists four fits: connected multi-shot scenes, self-hosted video pipelines where customisation, data locality or avoiding per-generation API fees matter, custom models built through fine-tuning and LoRA or IC-LoRA workflows, and developer integrations via the hosted API. For an app that needs dialogue, ambience or music matched to the picture, getting both from one model removes a separate audio-generation and syncing step.

The limits are just as concrete:

• Hardware. The claimed roughly 80 GB GPU for local use puts self-hosting out of reach for many independent developers without rented cloud GPUs.
• No independent benchmarks in the sources. The quality claims come from Lightricks, and the sources include no third-party leaderboard results for LTX-2.5.
• Output checks. The Rundown advises reviewing character continuity, prompt adherence, text, hands, motion artifacts and audio synchronisation before using a clip.
• Bias and factuality. The pre-trained model card says the model is not intended or able to provide factual information and may amplify existing societal biases. It adds that fine-tuned derivatives may introduce additional capabilities, limitations and safety risks and should be evaluated independently.
• The base checkpoint is not the product. Lightricks says the pre-trained checkpoint is not optimised the way the main LTX-2.5 generation model is.

The technical underpinnings are described in Lightricks’ paper “LTX-2: Efficient Joint Audio-Visual Foundation Model,” published on arXiv on January 6, which developers can use to understand the joint audio-video approach behind the 2.5 release.

What to watch next

Three things will decide how useful LTX-2.5 is outside well-funded studios. First, real memory figures: whether Lightricks or independent testers confirm the roughly 80 GB claim, and how far the quantized checkpoints in its collection can bring that down. Second, the maturity of the Diffusers integration, which The Rundown describes as early. Third, independent quality comparisons, which the sources do not yet provide. On the licensing side, how the community fine-tune rules play out as derivatives are shared will shape whether an ecosystem of LTX-2.5 LoRAs grows the way earlier LTX releases did.