NodeValidhttps://nodevalid.com/2026-09-12T00:00:00ZNodeValid: published video model weightshttps://nodevalid.com/#2026-09-122026-09-12T00:00:00ZWhich video model weights vendors have released, under what licence and on what hardware, and which version the hosted services sell instead.Wan: which versions have published weightshttps://nodevalid.com/models/wan/#2026-09-122026-09-12T00:00:00ZWhich Wan versions the vendor publishes weights for, under what licence, on what hardware, and which version the hosted services sell instead.LTX-Video: what Lightricks has publishedhttps://nodevalid.com/models/ltx-video/#2026-09-122026-09-12T00:00:00ZThe LTX-Video weights published under Apache 2.0, the variants released for lower VRAM, and the version sold only through the API.LTX Studio: sold version against published weightshttps://nodevalid.com/hosted/ltx-studio/#2026-09-122026-09-12T00:00:00ZThe model version LTX Studio sells by the second, whether weights for it are published, and which releases in the same family are open.SceneMixer: sold version against published weightshttps://nodevalid.com/hosted/scenemixer/#2026-09-122026-09-12T00:00:00ZThe model version SceneMixer sells by the second, whether weights for it are published, and which releases in the same family are open.What it takes to run the open weightshttps://nodevalid.com/requirements/#2026-09-122026-09-12T00:00:00ZThe hardware figures the Wan and LTX-Video repositories publish, why one repository gives two of them, and what a per-second rate replaces.What counts as published weightshttps://nodevalid.com/reading/#2026-09-122026-09-12T00:00:00ZWhy only vendor organisations count as a release, why the version sold gets its own row, and why hardware figures are quoted rather than derived.HunyuanVideo: what the repository publisheshttps://nodevalid.com/models/hunyuanvideo/#2026-09-122026-09-12T00:00:00ZThe licence, model size, stated hardware and resolution HunyuanVideo publishes in its own repository, with the date each was read.CogVideoX: what the repository publisheshttps://nodevalid.com/models/cogvideox/#2026-09-122026-09-12T00:00:00ZThe licence, model size, stated hardware and resolution CogVideoX publishes in its own repository, with the date each was read.Mochi 1: what the repository publisheshttps://nodevalid.com/models/mochi/#2026-09-122026-09-12T00:00:00ZThe licence, model size, stated hardware and resolution Mochi 1 publishes in its own repository, with the date each was read.Open weights come under two kinds of licencehttps://nodevalid.com/questions/two-kinds-of-open/#2026-09-122026-09-12T00:00:00ZThree families here carry a standard permissive licence. One carries a licence its vendor wrote, and one splits by model size.From four gigabytes to eightyhttps://nodevalid.com/questions/vram-ladder/#2026-09-122026-09-12T00:00:00ZPublished hardware figures across these families span a factor of twenty, from a laptop-class card to a rented data-centre one.Why naming the GPU beats stating a memory figurehttps://nodevalid.com/questions/naming-the-card/#2026-09-122026-09-12T00:00:00ZTwo families name the card their figures were measured on. That is the difference between a number a reader can price and one they have to interpret.How long a clip each family stateshttps://nodevalid.com/questions/length-across-families/#2026-09-122026-09-12T00:00:00ZPublished lengths run from a few seconds up to sixty, and the repositories state them in different units, so the table carries both.Published families that no hosted service nameshttps://nodevalid.com/questions/published-but-not-sold/#2026-09-122026-09-12T00:00:00ZTwo of the five families here have no counterpart on any priced page in this register, so their cost cannot be compared with anything.Apache 2.0 on both, and it settles less than it seemshttps://nodevalid.com/questions/licence-is-not-the-question/#2026-09-122026-09-12T00:00:00ZBoth open releases here carry the same permissive licence, so the licence cannot separate them; the version that is published is what does.The version sold is not the version publishedhttps://nodevalid.com/questions/which-version-is-open/#2026-09-122026-09-12T00:00:00ZWeights exist for two earlier generations of one family, while the generation that hosted services actually sell has none published at all.What the repository says you can run on one cardhttps://nodevalid.com/questions/what-you-can-run-at-home/#2026-09-122026-09-12T00:00:00ZA 5B model is described as running on a consumer graphics card, with a stated time for producing a five-second clip at 720P.Why one repository publishes two hardware answershttps://nodevalid.com/questions/the-two-hardware-numbers/#2026-09-122026-09-12T00:00:00ZThe same repository names a consumer card for one command and an 80GB floor for others, and both of those figures are correct.How long a clip the published weights can makehttps://nodevalid.com/questions/how-long-a-clip/#2026-09-122026-09-12T00:00:00ZOne repository states support for up to sixty seconds, far longer than any single generation the hosted services in this register sell.Renting seconds against running your own cardhttps://nodevalid.com/questions/buy-or-run/#2026-09-122026-09-12T00:00:00ZBoth hosted services here publish per-second rates, but comparing them with local compute means comparing two different versions of the model.What a model licence actually governshttps://nodevalid.com/learn/reading-a-model-licence/#2026-09-122026-09-12T00:00:00ZA permissive licence on published weights answers fewer questions than it appears to. These are the ones it answers and the ones it leaves entirely open.Reading a stated hardware requirementhttps://nodevalid.com/learn/planning-local-hardware/#2026-09-122026-09-12T00:00:00ZRepositories state hardware in their own terms, and two figures rarely describe one run. What a VRAM number covers, what moves it, and what it leaves out.