What open weights means, and what the licence lets you do

Open weights means you can download the model itself. The files are the trained parameters, often tens of gigabytes of them, and with them you can run the model on your own machine or on a rented one without asking anyone's permission for each clip. That is less than open source. The Open Source Initiative's definition (version 1.0, October 2024) asks for the parameters, the complete code used to train and run the system, and enough information about the training data that a skilled person could build a substantially equivalent model. None of the video models in this module publishes its training data, so none of them reaches that standard, whatever the launch post calls it.

It helps to see what arrives when you download one. A model's page on Hugging Face holds the weights as a set of large files, a model card explaining what the model does and how to run it, the configuration the code needs, and a licence file. Nothing runs by itself. You load the weights into software such as Hugging Face's diffusers library or the ComfyUI node editor, on a graphics card with enough memory to hold them, and the clip is made on that machine. With a hosted model such as Veo you rent access one clip at a time under the maker's terms of service, and the model never leaves the maker's servers. With open weights you hold a copy. It keeps working if a hosted service closes, as OpenAI's Sora API did on 24 September 2026, and unreleased product footage never has to leave a machine you chose.

The licence decides what you may do. In September 2026 the current models sit under four very different kinds of licence. Wan 2.2 from Alibaba is Apache 2.0, which allows commercial use anywhere with no revenue cap. LTX-2.5 from Lightricks uses the LTX-2.x Community License, free for any organisation with annual revenue under $10 million once its affiliates are counted in, and paid above that line. HunyuanVideo-1.5 from Tencent grants nothing in the European Union, the United Kingdom or South Korea, and asks any company with more than 100 million monthly active users to request a licence. MiniMax H3, downloaded 3.6 million times on Hugging Face in the month to 25 September, excludes those three places and the United States as well.

Two of those are standard licences. Apache 2.0 and MIT were written for ordinary software long before video models existed, and both let you use, change and sell what you make with the weights, for any purpose and in any country. Their duties begin when you pass the weights on to someone else. MIT asks you to keep its copyright and permission notice with every copy. Apache 2.0 asks for the licence text, the original notices and a note on any file you changed, gives you a licence to the contributors' patents that ends if you sue over them, and grants no right to the maker's trade marks. Neither says a word about disclosure, territory or revenue, which is why a studio can put an Apache model to work on paid jobs without asking anyone.

Community licences are longer, and each one is different, so read them in a fixed order. Start with the definitions, because defined words such as affiliates, derivatives and output decide how far every later clause reaches. Then find four clauses. The territory clause says where you may run the model and where its output may be shown. The threshold clause names the revenue or user count above which you need a paid or written licence, and the derivatives clause says what happens to a fine-tune or LoRA you make. The acceptable use policy, often a separate page, lists what you may not generate. Note the date and version on each.

Here is the client-work version. Wan 2.2, the older Mochi 1 preview (Apache 2.0) and Meituan's LongCat-Video (MIT) can be used for paid work anywhere, and the main duty is to keep the licence and its notices with any weights you pass on. LTX-2.5 suits a studio under $10 million a year, but its section 3.5 says a fine-tuned LTX model or LoRA handed to a client above that line needs the client to hold its own paid licence first. Tencent's and MiniMax's terms reach past where the studio sits, because both forbid displaying the output outside their territory, so a HunyuanVideo clip made in Mumbai still cannot run in London and an H3 clip cannot run in London or New York. H3 also needs written authorisation above $20 million of yearly revenue, and requires 'MiniMax H3' to be shown in the interface of any commercial product built on it. CogVideoX1.5-5B needs you to register with Zhipu before commercial use and caps you at a million visits a month.