Lightricks Gates Free Use of LTX-2.5 at $10 Million Revenue

Lightricks released LTX-2.5, a 22-billion-parameter open-weight video model, on Aug. 11 under a license that offers free commercial use only to companies under $10 million in annual revenue.

The model card, published on Hugging Face, names the LTX-2.x Community License Agreement as the governing terms. Revenue is measured "across the whole entity, including subsidiaries and affiliates under common control," according to the license, and companies above the threshold need a separate paid agreement with what Lightricks calls "full weights, engineering support, LoRAs, and flexible deployment options."

LTX-2.5's distilled version generates a 10-second, 720p clip in 6.8 seconds when self-hosted on two Nvidia GB200 superchips, according to cloud provider Runpod, which added the model to its own hosting options. Through Lightricks' own API, the same task takes 23.7 seconds end to end at 1080p.

What's new since LTX-2.3

Lightricks added four capabilities on top of the prior release, according to a ComfyUI Wiki summary of the model card: native multishot generation that holds character and lighting consistent across cuts, automatic duration prediction from the described action, native 4K HDR output, and a rendering mode Lightricks calls Diffusion Fidelity Rendering for cinema-grade EXR files.

The distilled model's weights run about 66 gigabytes on disk. Lightricks lists a 16GB VRAM floor for quantized versions and recommends 48GB or more for the full bf16 pipeline.

The speed claim is Lightricks' own

Runpod's write-up includes a same-task speed comparison Lightricks supplied: Gemini Omni Flash took 52 seconds, Grok 1.5 took 63 seconds, MiniMax H3 took 180 seconds, Seedance 2.5 took 317 seconds and Kling 3.0 Pro took 398 seconds, against LTX-2.5's 6.8 seconds.

1. LTX-2.56.8 seconds
2. Gemini Omni Flash52 seconds
3. Grok 1.563 seconds
4. MiniMax H3180 seconds
5. Seedance 2.5317 seconds
6. Kling 3.0 Pro398 seconds

Source: Lightricks benchmark, via Runpod, accessed 2026-09-12

None of those competing figures have been independently verified, and Lightricks ran every test itself. A separate comparison to Google's Veo 3.1, at 70 seconds, covers an 8-second clip rather than the 10-second standard used for the rest, making it the one figure in the set that is not directly comparable.

The license terms put LTX-2.5 on different ground than other open-weight releases this month, including DeepSeek's V4.1-Flash and OpenBMB's MiniCPM5-2B, both released under a plain MIT or Apache license with no revenue threshold attached.

Lightricks' model card also references a full, non-distilled "dev" version of LTX-2.5 without giving a release date or saying how its license will differ from the distilled model's terms.

The Hugging Face repository itself is gated: downloading the weights requires accepting the license on the page first, rather than a plain click-through download. That puts a paper trail behind every download, which is also how Lightricks would identify a user who crossed the $10 million threshold without moving to the paid agreement.

For a small studio or an agency under that revenue line, the practical effect is a video model with delivery speeds close to Lightricks' own hosted API, running on hardware the team owns. For anyone bigger, the 6.8-second number is a demonstration of what the paid tier buys rather than a price they can use for free.