The weights that are open are not the ones being sold

NodeValid records which video model weights a vendor has actually released, under what licence and on what hardware. Two of the 15 families here have a counterpart on a priced page, and in both of those the published weights are an earlier generation than the version billed by the second. As of 2026-09-22.

Inclusion rule. Model families whose vendor has published at least one release openly, read from an organisation that vendor controls. Where no hosted service in this register sells a version of the family, the last two columns are recorded as not applicable. Order. Alphabetical by family.

1How this register is laid out

Model pages and hosted pages, kept apart. A model page records what a vendor has released from its own organisations: versions, licence, and any hardware figure the repository itself states. A hosted page records which version a paid service sells and whether weights for that exact version exist in the open.

Keeping them apart matters because the brand is shared and the version numbers are close. Wan 2.2 is open; Wan 3.0 is sold. LTX-Video is open; LTX-2.5 is sold. Reading either pair as one thing turns every cost and quality comparison into a comparison of different models.

2Start here

• Wan — 2.1 and 2.2 published under Apache 2.0; 3.0 is not.
• LTX-Video — 13B with distilled and fp8 variants, Apache 2.0.
• HunyuanVideo — 13B under a licence the vendor wrote itself.
• CogVideoX — one family, two licences, and a 4GB entry point.
• Mochi 1 — Apache 2.0, 10B, an H100 and 480p.
• Allegro — two memory figures for one run, 9.3G and 27.5G.
• Step-Video-T2V — 30B under MIT, with a memory table of three rows.
• Open-Sora — 11B, and the memory figure moves with the card count.
• Stable Video Diffusion — a community licence that routes commercial use elsewhere.
• FramePack — 6GB, and the figure says what 6GB produces.
• Pyramid Flow — precise about output, silent about size and memory.
• SkyReels-V2 — 14.7GB or 51.2GB for the same 540P frame.
• MAGI-1 — hardware written as eight H100s or one 4090.
• Cosmos Predict 2.5 — every field filled, down to two decimal places.
• LongCat-Video — MIT and 13.6B, with no memory figure at all.
• SceneMixer — sells Wan 3.0 seconds, names the version on its price list.
• LTX Studio — sells LTX-2.5 seconds while its parent publishes LTX-Video.
• Requirements — a 4090 and 80GB of VRAM, both from the same repository.
• Method — vendor organisations only, hardware quoted not derived.
• Open or hosted — four searches about open weights, answered from the register's own rows.
• Two kinds of open — open weights come under two kinds of licence
• The VRAM ladder — from four gigabytes to eighty
• Naming the card — why naming the GPU beats stating a memory figure
• How long a clip — how long a clip each family states
• Nobody sells these — published families that no hosted service names

3Questions the register answers

• Two kinds of open — open weights come under two kinds of licence
• The VRAM ladder — from four gigabytes to eighty
• Naming the card — why naming the GPU beats stating a memory figure
• How long a clip — how long a clip each family states
• Nobody sells these — published families that no hosted service names
• Licence vs weights — Apache 2.0 on both, and it settles less than it seems
• Which version — the version sold is not the version published
• Running it locally — what the repository says you can run on one card
• 4090 or 80GB — why one repository publishes two hardware answers
• Clip length — how long a clip the published weights can make
• Buy or run — renting seconds against running your own card
• Same licence, different reach — one licence line, seven different answers
• Floor or peak — which repositories state a minimum rather than a peak
• Figures, no card — which releases give gigabytes and name no card
• Frames, no rate — which releases publish frames and no frame rate
• Exact or rounded — which vendors publish an exact parameter count
• Complete output — which releases state size, rate and length together
• Two figures, one run — which repositories publish more than one memory figure
• Which variant — which families attach their figures to a named variant
• Consumer cards — which releases name consumer hardware at all
• Stated settings — which settings are published beside a memory figure
• Card or repository — where the licence value is actually published
• More than one card — which releases document running across several cards
• Empty cells — which cells of the register are empty, and why

4Explainers

Craft notes on reading releases: what a licence file governs, and how to read a stated hardware figure. These pages carry no vendor facts of their own — the sourced rows are on the model and hosted pages.

• Reading a licence — what a licence on published weights governs, and what it leaves untouched.
• Planning hardware — how to read a stated VRAM figure, and what it leaves out.

5Why a parameter count is not a requirement

The most repeated number about any open model is its parameter count, and it is the one that answers the fewest questions. A 5B model and a 14B model from the same family can need a consumer card and 80GB of VRAM respectively, and the repository says so in both cases. Two entries here publish around thirteen billion parameters and state 60GB and 6GB of memory respectively, which settles the question of whether a count predicts a requirement. This register quotes the statements and estimates nothing from size.

The licence is the other fact everybody repeats, and it travels better as long as it is read as a name rather than as a category. Eight distinct licence names sit across the families here: two standard ones a reader clears once and for all, five that carry a company's name inside the licence itself, and one that changes with the model size inside a single project. None of them describe the versions being sold, which carry no public licence because they carry no public weights.

NodeValid records published releases and published rates. It does not benchmark models and it does not recommend running anything locally.

6Reading the register by section

The weights and hardware sections are read from repositories; the hosted rows are read from price pages. Between them: what a licence covers, and where released weights differ from the service on sale.

• Data — Recorded licences, sizes and hardware notes for open video-model weights…
• Models — An entry per weight family…
• Hosted — Vendors that sell access to a model while publishing weights for an earlier version…
• Questions — Questions about running video models locally, answered from repositories, licence files…
• Learn — Explainers on reading open-weight releases…
• Fields — The columns this register reads every weight release into…
• Statements — One page per family per register column…
• Contrasts — Pairs of weight families that answer one register column differently
• Terms — One note per word a model card uses and never defines
• Open or hosted — Four pages on the gap between a weights page and a price page…