LTX-2.5's Free Commercial Licence Stops at $10 Million in Group Revenue, and Its Card Asks for Contact Details
A.I. / news
LTX-2.5's Free Commercial Licence Stops at $10 Million in Group Revenue, and Its Card Asks for Contact Details
Lightricks' open-weight video and audio model is free for production use below that line, but the revenue test counts parent companies and affiliates, and the card publishes no benchmark scores.
Lightricks' LTX-2.5, an open-weight model that generates video with synchronized audio, is free for commercial use only to organisations under $10 million in annual revenue. Above that, its Hugging Face model card says a paid Commercial Use Agreement is required.
The weights are on Hugging Face's trending list with 6,849 likes and about 1.69 million downloads. The card gives no release date. Its licence, the LTX-2.x Community License Agreement, is dated Aug. 11, 2026, according to the Hakky Handbook's licensing guide, last updated Oct. 1.
How the $10 million test works
The revenue measure is the whole entity, not the project. The card says it includes subsidiaries and affiliates under common control. By that wording, a startup with $2 million in sales that sits inside a group with $40 million would not qualify for the free tier.
Under the line, the card allows commercial and production use at no cost, and Hakky says that covers self-hosting, fine-tuning and deployment. At or above it, the paid agreement includes full weights, engineering support and LoRAs, which are small add-on adapters that steer a model toward a style or subject. The licensing contact in the guide is [email protected].
One line in the card is easy to miss: transferring fine-tunes may require a paid licence. Hakky describes redistribution of the base model or fine-tunes as "subject to separate conditions" and says anyone who passes on copies "must include the license and Attachment A". That attachment holds the acceptable-use rules. Neither source lists them, so a team should read the text in the repository before building on it.
| Organisation size | Commercial use | What the card or guide says |
|---|---|---|
| Under $10 million annual revenue | Free | Production use permitted |
| $10 million or more | Paid agreement | Includes full weights, support, LoRAs |
| Any, redistributing fine-tunes | Conditions apply | Recipient must be bound by the same terms |
What the model is and what it needs
The card describes text-to-video, image-to-video and video-to-video generation, with audio produced through a separate audio VAE and vocoder. It does not state a parameter count. File names for the diffusion transformer carry "22b", which suggests 22 billion, but that is an inference from names.
There are two variants: a distilled one that runs a fixed 8-step schedule with guidance set to 1, and a full dev version meant for training. Formats are bf16, int8 (ComfyUI only) and NVFP4, which targets Nvidia's Blackwell GPUs. The card gives no VRAM figure and recommends Python 3.12 or newer and CUDA 12.7 or newer.
Width and height must divide by 32, and frame counts must satisfy frames % 8 == 1. The card's example renders 121 frames at 24 frames per second, about five seconds, and reaches 1920 by 1088 or, with an extra upscale pass, 3840 by 2176.
What the card does not say
There are no quantitative results. The card describes improvements over LTX-2 only in words: better detail, prompt adherence and multishot generation. It lists limits too: the model cannot provide factual information, may amplify bias, reacts strongly to prompt style and can produce inappropriate content. The paper it links, arXiv 2601.03233, is listed as published Jan. 6.
Access is also not as open as the phrase suggests. As fetched, the card asks users to agree to share contact information before download. It is a community licence with a revenue cap, a different thing from the Apache 2.0 terms on Qwen3.8-27B, whose largest sibling carries its own $50 million threshold.
A model priced by licence, not by token, also changes the comparison with hosted services such as Claude Haiku 5.5 at $0.10 per million input tokens: the cost here is GPU time and, above $10 million, a negotiated contract with no published price.
The figure to watch is that contract. Lightricks lists none, and the card has no date for when LTX-2.5 was released.
Sources
More in A.I.
- 01Gemini 4 Argon Costs $1.99 a Task at Promo Price Against $0.72 for GPT-6.1 Sol, and Is Not Yet on SaleGoogle's introductory $2 and $10 rates match OpenAI's Sol per token, but Artificial Analysis figures cited by eesel show Argon writing 62,000 output tokens a task where GPT-6 Astra writes 27,000.
- 02EmbeddingGemma 2 Embeds Text, Images, Audio and Video in 567MB of RAM on a Pixel 11 ProGoogle DeepMind's Apache 2.0 embedding model has 740M parameters in three modular pieces, and the only benchmark number its launch post prints is a 9.92-point gain on MTEB Code.
- 03Kolibri-1 Is Apache 2.0 and Fits on One B200, but Trails Qwen3.8 27B by 9.1 Points in GermanAleph Alpha's 78B-parameter mixture-of-experts model activates 3.46B parameters per token, and its own model card shows a larger dense Qwen ahead on every headline benchmark.
- 04Qwen3.8-27B Is Apache 2.0 With a 262,144-Token Window and Already Has 482 Finetunes, Including Cloudflare's ClefAlibaba's dense 27B model claims 61.7 on SWE-bench Pro against 53.4 for Opus 4.6 Max, and the model card says nothing about training data.