What Is LTX-2.5? Lightricks Open-Weights Video Model
Architecture, Licensing Boundaries, API Pricing, and Local Deployment Options
LTX-2.5 is an open-weights 22-billion-parameter (22B) asymmetric dual-stream diffusion transformer model developed by Lightricks for joint audio and video generation. Released on the Lightricks website in October 2026, it succeeds LTX-2.3 with native multi-shot sequencing, automatic duration prediction, and diffusion fidelity rendering.
The model includes public weights on Hugging Face and source code on GitHub. Lightricks terms permit free commercial use for organizations generating under $10 million in revenue or Annual Recurring Revenue (ARR), while higher-revenue organizations require a custom commercial license.
At Layer3Labs, we help organizations map and integrate automation infrastructure into production workflows, where model licensing terms, compute requirements, and integration surfaces determine operational feasibility. Assessing LTX-2.5 requires reviewing its dual-stream architecture, calculating Application Programming Interface (API) generation expenses, and auditing technical specifications that remain unpublished.
What LTX-2.5 Is and Who Can Use It
LTX-2.5 is a 22B-parameter open-world video model created by Lightricks that generates synchronized video and audio from text prompts, static images, or audio reference files. Lightricks publishes both the model weights and underlying code, categorizing the architecture as an asymmetric dual-stream diffusion transformer.
The model targets production video generation by producing cinema-grade visual details, maintaining consistent character features across multi-shot sequences, and rendering legible on-screen text. Lightricks describes the core mechanics as bidirectional cross-attention paired with modality-aware classifier-free guidance, documented in the technical report on arXiv.
Commercial eligibility depends on company revenue. Lightricks allows organizations generating under $10 million in annual revenue to deploy LTX-2.5 without license fees or mandatory branding, while larger commercial deployments require custom agreements negotiated through the Lightricks model page.
- Developer: Lightricks
- Parameter count: 22 billion parameters across an asymmetric dual-stream architecture
- Modality: Native joint audio and video generation
- Asset availability: Model weights on Hugging Face and code repositories on GitHub
- Hugging Face repository license tag: Classified as 'other' with automated access approval
- Commercial threshold: Free commercial rights for companies under $10 million in revenue or ARR per Lightricks statements
- Deployment methods: Cloud API endpoints, on-premises infrastructure, local workstation software, and browser interfaces
New Architecture and Changes from LTX-2.3
LTX-2.5 introduces structural upgrades over LTX-2.3, including an updated diffusion video decoder, an integrated Gemma 4 12-billion-parameter (12B) text encoder, and reinforcement learning (RL) training on an expanded dataset. Lightricks lists cleaner motion among the results, with fewer motion artifacts and smoother, more natural output.
A primary feature is Diffusion Fidelity Rendering, which allocates rendering compute according to scene complexity. The model synthesizes frames from a grid of high-fidelity keyframes, delivering pixel quality that holds up frame by frame.
Creative prompting is managed by the Gemma 4 12B encoder combined with an automated prompt enhancer. Lightricks reports that this combination improves creative instruction following from shorter text descriptions, while automatic duration prediction calculates the output clip length directly from the prompt action.
- Diffusion Fidelity Rendering: Rendering compute allocated by scene complexity
- Native multi-shot: Wide, medium, and close-up camera angles within a single generation pass while preserving lighting, environment, and character identity
- Cleaner motion: Fewer motion artifacts and smoother, more natural results
- Prompt adherence: Shorter input prompts parsed by a Gemma 4 12B text encoder and custom prompt enhancer
- Automatic duration prediction: Clip duration calculated from prompt action rather than manual duration sliders
- Architectural updates: New diffusion video decoder, updated distilled model, and RL optimization
Post-Production Tools and Foundation Fine-Tuning
Lightricks packages LTX-2.5 with specialized post-production utilities and a downloadable pretrained checkpoint intended for domain-specific fine-tuning. The pretrained checkpoint allows teams to perform deep fine-tuning on their own data and domain.
Several of the post-production tools, namely Refine, Restore, the upgraded SDR to HDR and Layout to Render, are offered as IC-LoRA downloads on the LTX-2.5 product page.
Color workflows support high-end finishing pipelines. The model processes 16-bit OpenEXR (EXR) source files through its HDR Native module and outputs native High Dynamic Range (HDR) video, with the Lightricks model page citing 4K HDR output.
- Native Resolution: Enables full delivery resolution generation on a single graphics processing unit (GPU) with minimal Video Random Access Memory (VRAM) expansion
- Refine: Sharpening utility using IC-LoRA to restore blurry or pixelated footage
- Restore: Archival footage restoration using IC-LoRA modules to add color
- SDR to HDR: Format upgrade tool converting Standard Dynamic Range (SDR) video into HDR delivery files
- HDR Native: Ingests 16-bit EXR input files to deliver native cinema-grade HDR output
- Layout to Render: Three-dimensional (3D) camera-aligned background generation using spatial blocking and first-frame guidance
- AlphaGen: Beta utility for creating soft alpha mattes for compositing workflows
Deployment Options across API, Local Hardware, and Apps
Organizations can deploy LTX-2.5 through managed cloud endpoints, local application environments, or direct self-hosting on private GPU infrastructure. Weights are hosted in a gated repository on Hugging Face with automatic access approval, tagged for image-to-video, text-to-video, audio-to-video, and video-to-audio pipelines across nine languages.
Self-hosted environments run code from GitHub. Teams planning local workstation deployments can assess GPU capacity requirements with our Local AI Hardware Calculator, while creative pipelines migrating from earlier models can review our guide on LTX-2 for creative teams.
Desktop operators can use LTX Desktop, which runs the full video editing environment locally on workstation hardware. Hosted access is available through the LTX Playground and subscription plans on the LTX Studio pricing page, which list Free, Lite ($15 monthly), Standard ($35 monthly with commercial rights), and Pro ($125 monthly) tiers.
- Hugging Face repository: Gated repository at Lightricks/LTX-2.5 with automated access approval, last modified in October 2026
- GitHub repository: Open-source implementation files maintained at Lightricks/LTX-2
- LTX Desktop: Local video editor powered directly by the LTX-2.5 engine on local workstation GPUs
- LTX Studio: Hosted cloud software offering commercial plans starting at $35 monthly for business use
API Pricing and Cost Estimates for LTX-2.5
The LTX API bills video generation on a per-second basis, with no request fees or per-asset charges. Rates published on the LTX API pricing page depend on the model variant, output resolution, and generated clip length, with portrait and landscape formats priced identically.
The API offers two model variants: LTX-2.5 Fast for rapid previews and creative exploration, and LTX-2.5 Pro for production renders requiring higher temporal stability. Both variants support resolutions from 720p through 4K across synchronous and asynchronous endpoints.
Audio-to-video generation accepts WAV, MP3, M4A, and OGG audio inputs up to approximately 20 seconds per request, billed at the same per-second video rates. For the full cost breakdown, see our guide on LTX-2.5 pricing and our general AI model pricing index.
- LTX-2.5 Fast API rates: 720p ($0.09 per second), 1080p ($0.13 per second), 1440p ($0.19 per second), 4K ($0.30 per second)
- LTX-2.5 Pro API rates: 720p ($0.12 per second), 1080p ($0.17 per second), 1440p ($0.25 per second), 4K ($0.39 per second)
- 10-second 1080p Fast clip: 10 seconds multiplied by $0.13 equals $1.30
- 10-second 1080p Pro clip: 10 seconds multiplied by $0.17 equals $1.70, reflecting approximately 31 percent higher cost than Fast
- 60-second 1080p Pro total: 60 seconds multiplied by $0.17 equals $10.20
- 100 iterations of 10-second 1080p Fast previews: 1,000 seconds multiplied by $0.13 equals $130.00
- Comparison with LTX-2.3: At 1080p, LTX-2.5 Pro ($0.17) is $0.09 per second above LTX-2.3 Pro ($0.08), while LTX-2.5 Fast ($0.13) is $0.07 above LTX-2.3 Fast ($0.06)
Commercial License Terms for LTX-2.5
Verifying commercial usage rights requires reviewing the exact license file on Hugging Face because Lightricks publishes differing definitions across its web properties. The primary overview on the Lightricks model overview states the model is free for companies under $10 million in revenue, while the release page on the LTX-2.5 product page specifies under $10 million in ARR.
The Hugging Face repository labels the license tag as 'other', which is not a standard SPDX or permissive license. Access to the checkpoints on Hugging Face is gated, with automatic approval.
Commercial organizations surpassing the $10 million threshold in annual revenue or ARR cannot rely on the free tier. Lightricks offers a commercial license for teams building at scale.
Unpublished Specifications and Pre-Deployment Testing
Lightricks leaves several operational specifications unpublished, requiring technical teams to benchmark the model directly on local hardware. No benchmark scores for LTX-2.5 are given in the vendor pages reviewed for this guide.
Hardware requirements are only partly stated. A comparison table on the LTX-2.5 product page lists a "Min VRAM: 16GB" row, but it is a marketing comparison, not a hardware specification, and no tested inference speed is given in the sections reviewed, so verify on the vendor page and test on your own hardware. Furthermore, maximum clip length limits for the raw open weights and the exact credit-to-second conversion rates in LTX Studio are not published.
Teams evaluating LTX-2.5 should clone the repository from GitHub and execute proof-of-concept tests on target hardware before committing production resources.
What you need to run LTX-2.5 yourself
LTX-2.5 needs real memory, but it is within reach of a high-end workstation or a couple of professional GPUs — and many teams simply rent instead of buying. Match the path below to whether you want to own the hardware or pay by the hour.
| Path | What it is | Best for | Get started |
|---|---|---|---|
| Call the hosted API | Use LTX-2.5 as a pay-per-token API — zero hardware | Most teams; getting started | OpenRouter |
| Rent GPUs by the hour | Spin up an H100 / A100 for a few dollars an hour | Flexible self-hosting without buying cards | RunPod |
| Local on unified memory | One Mac with enough unified memory to hold a 4-bit quant | A single quiet on-prem box | Apple Mac Studio (M4 Max, 128GB) |
| Local on a workstation GPU | One 48GB pro card, or two 24GB consumer cards | Power users who want hardware they own | NVIDIA RTX 6000 Ada (48GB) |
To put LTX-2.5 to work once it is live, connect a coding client like Cursor (via OpenRouter) or a local runner such as Ollama.
Frequently Asked Questions
- LTX-2.5 is an open-weights video model developed by Lightricks featuring a 22B-parameter asymmetric dual-stream diffusion transformer for joint audio and video generation. It supports text-to-video, image-to-video, and audio-to-video creation with native multi-shot camera sequencing.
- Lightricks describes LTX-2.5 as open source with code on GitHub and weights on Hugging Face, but the repository carries a license tag marked 'other', which is not a standard permissive license.
- LTX-2.5 is free for commercial use by businesses generating under $10 million in revenue or ARR per Lightricks documentation, while larger companies must purchase a commercial license.
- LTX-2.5 natively generates synchronized audio and video together through bidirectional cross-attention, and it accepts external WAV, MP3, M4A, or OGG files for audio-driven video generation.
- LTX-2.5 adds native multi-shot generation, automatic duration prediction, Diffusion Fidelity Rendering, cleaner motion, a Gemma 4 12B text encoder, and an updated diffusion video decoder compared to LTX-2.3.
- You can run LTX-2.5 via the official Lightricks API, download open weights from Hugging Face for local or on-premises GPU hosting, execute it through LTX Desktop, or use hosted interfaces like LTX Studio and LTX Playground.
Plan Your Enterprise Video Automation Architecture
Speak with our implementation team to evaluate how open-weights video models like LTX-2.5 fit into your infrastructure, compliance requirements, and budget constraints.
Book a Consultation