Run the model on your own hardware with LTX-2.5.
LTX-2.5 is Lightricks’ current video model, and it is sold on a different axis from everything else here: not the longest clip or the highest resolution, but weights you can download, fine-tune and run on your own hardware.
Capabilities and specifications on this page come from Lightricks’ own LTX-2.5 model page, its API pricing documentation and the published licence terms. Prices and licence thresholds are set by the vendor and change without notice.
What LTX-2.5 renders
These frames come from Lightricks’ own LTX-2.5 model page. The vendor publishes them as capability demonstrations rather than as a labelled gallery, so the captions below describe what is in each frame and make no claim about which setting produced it.
Example images belong to Lightricks and are reproduced with credit: LTX-2.5 model page
How Lightricks describes LTX-2.5
Lightricks markets 2.5 as a foundation model — something a studio builds on — rather than as a hosted generator. Its own model page argues that with eight improvements and one comparison table.
Open weights is the argument
Lightricks’ own comparison table lines up the requirement rows — downloadable weights, fine-tuning on your own data, runs on any GPU, on-prem or edge deployment — against MiniMax H3, Seedance 2.5, Kling 3.0 Pro and Veo 3.1, and marks those four partial or closed on the same rows.
Diffusion Fidelity Rendering
The name Lightricks gives its new rendering approach: compute is allocated by scene complexity rather than spread evenly, which the vendor says is why quality holds up frame by frame on a cinema screen.
Multi-shot continuity
One prompt can direct several connected cuts, holding character, environment, lighting and voice across them — the thing single-shot generators cannot do.
Built to be fine-tuned
Alongside the production model, Lightricks ships a pretrained foundation checkpoint that has not been through supervised fine-tuning, aimed at teams taking a video model into robotics and industrial simulation.
Duration decides itself
The model predicts a clip length from the requested action instead of asking you for a number. LTX-2.5 ships with a new diffusion video decoder, a Gemma 4 12B text encoder and a custom prompt enhancer.
What the official API charges per second
Text-to-video and image-to-video bill by the second of output, and the audio track costs nothing extra at any resolution. Audio-to-video bills on the length of the track you upload instead. Pro has no 1440p or 4K tier in the current API.
| Tier | Resolution | List price |
|---|---|---|
| LTX-2.5 Fast | 720p | $0.09 / second |
| LTX-2.5 Fast | 1080p | $0.13 / second |
| LTX-2.5 Fast | 1440p | $0.19 / second |
| LTX-2.5 Fast | 4K | $0.30 / second |
| LTX-2.5 Pro | 720p | $0.12 / second |
| LTX-2.5 Pro | 1080p | $0.17 / second |
Ten seconds of 1080p costs $1.30 on Fast and $1.70 on Pro. Because the model decides the length itself, you pay for the seconds it actually produces. Lightricks notes on its own page that the speed and artifact comparisons it publishes are vendor-run and use different resolutions for some competitors.
The eight things LTX-2.5 changes
This is Lightricks’ own capability list for the release, in the vendor’s order and wording.
Cleaner motion
Fewer motion artifacts, and smoother, more natural movement.
Multishot
Connected shots that hold character, environment, lighting and voice.
Better prompt adherence
Complex creative instructions followed from shorter, simpler prompts.
Auto duration
The model generates the right clip length for the action you asked for.
Native 4K HDR
High-resolution HDR footage built for professional finishing.
Native RAW workflow
Generate and edit inside professional colour and finishing pipelines without compromising the master.
Better results, less compute
Near-full-model quality with fewer retries, fewer controls and significantly lower compute.
A foundation you can fine-tune
Fine-tune a raw foundation model on your own data and domain.
Documented specifications
Every row below is stated on Lightricks’ LTX-2.5 model page or in its API documentation.
- Developer
- Lightricks
- Current model
- LTX-2.5
- Tiers
- LTX-2.5 Fast · LTX-2.5 Pro
- Weights
- Open
- Output
- Up to native 4K HDR
- Colour
- Native RAW workflow
- Duration
- Predicted automatically
- Text encoder
- Gemma 4 12B with a custom prompt enhancer
- Video decoder
- New diffusion video decoder
- Minimum VRAM
- 16 GB
- Deployment
- On-prem, edge or API
- Licence
- LTX-2.x Community Licence
Documented access channels
There are three routes to LTX-2.5 and they are priced differently: downloadable weights, the hosted API, and the company’s own studio product. The licence matters more here than on a closed model, because the weights are the point.
LTX-2.5 model page
Lightricks’ own page for the model: the capability list, the comparison table and the notes on how its speed figures were measured.
OpenAPI pricing
The official per-second rates for the Fast and Pro tiers, including the audio-to-video endpoint.
OpenWeights on Hugging Face
LTX-2.5 is published as Lightricks/LTX-2.5. The download is gated behind contact details, and it is governed by the LTX-2.x Community Licence rather than an OSI licence.
OpenOpen-source repository and tooling
Lightricks’ LTX-2 repository carries the code and the LoRA training tooling; the platform already lists more than 80 community fine-tunes of the LTX 2.x line.
OpenThird-party API providers
The hosted API is also served through third-party providers. The rates above are the official ones; a provider may differ, and its terms are its own.
LTX-2.5 questions
Is LTX-2.5 really open source?
The weights are downloadable, but the licence is not an OSI licence. LTX-2.5 ships under the LTX-2.x Community Licence, which allows local and production use and requires a paid agreement once an organisation passes a revenue threshold. Lightricks links the binding terms from its own pages — read those, not a summary.
What do I need to run LTX-2.5 locally?
Lightricks puts the minimum at 16 GB of VRAM, and the model is large enough that storage and setup matter too. Quantised and offload options exist, and the company says it runs on any GPU above that floor.
Can I fine-tune it?
Yes, and that is the argument for open weights a closed model cannot make. LoRA training adds a style, a character or a behaviour the base model never learned, and Lightricks also ships a pretrained checkpoint that has not been through supervised fine-tuning, aimed at domain adaptation.
Does LTX-2.5 generate audio?
Yes. Audio is produced with the picture and costs nothing extra at any resolution, and there is an audio-to-video endpoint that drives the clip from a track you supply — billed on the length of that track.
What does the API cost?
On the official rate card: Fast is $0.09 per second at 720p, $0.13 at 1080p, $0.19 at 1440p and $0.30 at 4K; Pro is $0.12 at 720p and $0.17 at 1080p, with no higher tiers in the current API. Ten seconds of 1080p is $1.30 on Fast and $1.70 on Pro.
Is LTX-2.5 faster than the closed models?
That is Lightricks’ claim, and the page publishes how it was measured: 6.8 seconds for a 10-second clip on its own dual-GB200 hardware, against 398 seconds for Kling 3.0 Pro in the same table. The API figure is 23.7 seconds, measured through a third-party provider with queue time included. The vendor notes that competitors were measured at different resolutions.
Does 2.5 have the editing tools earlier versions had?
Not all of them yet. The published documentation records that retake, extend and reframe remain on the LTX-2.3 endpoints rather than moving to 2.5, so a workflow that depends on them should stay on 2.3 for now.