LTX-2.5
A 22B-parameter diffusion transformer that generates synchronized video and audio in a single model, using a Gemma-class 12B text encoder and separate video and audio VAEs.
- Weights
- Downloadable
- API
- No
- Commercial use
- Permitted subject to conditions
Shipped as a full 'dev' transformer for guided two-stage pipelines plus a distilled variant that runs in far fewer steps. Lightricks publishes weights, inference code and training code.
LTX-2.x Community License Agreement, not an OSI licence. Free for commercial use below $10M annual revenue.
Revenue threshold: free below $10,000,000 in annual revenue, chargeable above it.
- Free only below a revenue threshold. Free for academic research and for commercial use by entities under $10M annual revenue.
Specification
- What it takes in
- text, image
- What it makes
- video, audio
- Maximum duration
- not established
- Maximum resolution
- not established
- Aspect ratios
- not established
- Controls
- text prompt, reference image, audio generation toggle, distilled 8-sigma fast pipeline, two-stage guided pipeline, FP8 quantization, CPU/disk offload, latent upscaling
- Weights
- Downloadable
- API
- No
- Pricing
- Free open weights under the LTX-2.x Community License for entities under $10M annual revenue. Self-hosting is heavy: the main bf16 checkpoint is about 43 GiB and the full component set (transformer, text encoder, video and audio VAEs, upscalers) is roughly 66 GiB, so a 48GB+ GPU or FP8 quantization plus offloading is needed. apiAvailable is set false because no first-party documented developer API endpoint for the open LTX-2.5 weights was verified; Lightricks exposes the technology through its own LTX Studio consumer product, and programmatic access is generally via third-party hosts (fal, Replicate, WaveSpeed), which does not count as a first-party API.
- Released or updated
- 2026-01-06
- Prompt limit
- not published by the vendorVendor figure, checked 2026-08-16Collected from the vendor's own documentation on 16 August 2026. Vendors change these between model versions without announcing it, so check the vendor's current API reference before you rely on it. Check the vendor's documentation
Maximum duration, maximum resolution and supported aspect ratios are not established here. That means no page we could retrieve stated them, not that the model lacks them.
Strengths
- Native synchronized audio plus video in one open-weights model, still rare
- Full training code published, not just inference
- Free commercial use below $10M revenue
- Distilled variant for fast inference
- Training data licensed from stock providers
Weaknesses
- $10M revenue threshold means large companies must buy a license
- Acceptable-use policy and anti-competing-model clause make it non-OSI
- Official docs do not publish max resolution, fps or duration numerically, so those fields are null
- 43-66 GiB of weights is a real deployment burden
Best for
- Startups and independent studios under $10M revenue that need open-weights video with synchronized audio and are willing to run large GPUs.
Where the sources disagree
Recorded rather than resolved. Picking the tidier number would hide the fact that the vendor's own pages do not agree.
- Marketing and third-party sources widely claim 'native 4K at 50 fps' and clip lengths of 6/8/10 seconds (or 20 seconds for a Pro tier), but neither the GitHub README nor the Hugging Face model cards state resolution, fps or duration limits numerically. Parameter count is also reported inconsistently: press coverage said 19B (14B video + 5B audio) for the January LTX-2 launch while the LTX-2.3/2.5 checkpoints are named 22b. Resolution, fps and duration left null.
What to understand before using it
Explainers naming the mechanisms this model's own documentation mentions.
MCP servers that front Lightricks
Servers in the media MCP directory whose own description or homepage names this vendor. Absence of a server here is not evidence that none exists: most publish no tool list at all.
-
Comfyui
MCP server + Claude Code plugin for ComfyUI: run workflows, generate images, manage models & VRAM.
-
Corent
Generate images and videos via the Corent API, auto model routing with provider fallback.
-
Fal
Create and edit high-quality media across various formats including images, videos, audio, and 3D models.
-
OpenArt
Generate AI images and video from any agent, with OAuth and no API keys.
-
Picsart GenAI
Generate and edit images, videos and audio with 150+ AI models.
-
Runcomfy
Remote MCP server for RunComfy Serverless API (ComfyUI): deployments and async inference.
See it from the vendor
Shotcraft hosts no sample output. Judging a model means watching its own showreel with your own eyes, and a clip re-hosted here would be a clip we chose.
Nearby in the directory
LTX-Video 13B 0.9.8 (and distilled / 2B variants)
The first-generation LTX-Video family: a 13B DiT plus a distilled 13B and a 2B distilled model, all built for speed via multiscale rendering. It is on
Video generation
Adobe Firefly Video Model
The Adobe Firefly Video Model generates video from a text prompt, optionally guided by an image used as the first or final frame. Adobe's Firefly Serv
Video generation
Allegro
A small text-to-video model comprising a 2.8B VideoDiT and a 175M VideoVAE, released under Apache-2.0 in October 2024. It generates 88-frame clips at
Video generation