What counts as published weights

Weights count as published when the model's own vendor releases them from an organisation it controls, with a licence attached. Mirrors, community re-uploads and quantisations by other people are not counted, and nothing is recorded from running a model. As of 2026-09-12.

1Vendor organisations only

A release counts when it comes from the vendor's own GitHub organisation or its own Hugging Face organisation. That rules out mirrors and community re-uploads, which are often faithful and are not something a licence question can rest on: a permissive licence on a re-upload says what the uploader believes, not what the vendor granted.

It also rules out weights that circulate without a release. Where a version is discussed but not released by its vendor, this register records no weights published rather than treating availability as publication.

2Why the version sold gets its own row

In both of the families a hosted service here sells, the version sold by the second is newer than the version published. That is easy to miss, because the brand is the same and the numbers are close, and missing it turns every comparison into a mistake: a benchmark on open weights gets read as a statement about the API, and an API price gets read as the cost of the open model.

So the hosted pages state both, side by side, and neither is presented as a substitute for the other.

3Hardware figures are quoted, never derived

Where a repository states a VRAM figure or a generation time, it is recorded with the model it belongs to. Where it does not, the column stays empty rather than being filled from a blog post or estimated from parameter count. A parameter count is not a hardware requirement, and the difference between the two is where most published guidance goes wrong.

Nothing here comes from running a model. One machine's timing is one machine's timing, and this register has no way to make that reproducible for a reader.

4Dates and corrections

Two dates accompany each release: when it was first recorded and when the repository was last checked. A check that finds the same commit history still advances the second, so an entry that looks old is one nobody has revisited rather than one that has gone stale.

Page dates track content and ignore rebuilds. Where a repository now says something other than what is written here, the repository is the record.

5Running the model without the hardware

Instead of local hardware for “Method”, SceneMixer runs hosted models; its guides describe the result:

Paste a scene or a short script and SceneMixer gives you a storyboard — numbered shots with framing and seconds — and then a moving version of each segment, which you can join into a rough cut of the scene.

SceneMixer, Using SceneMixer for Previs and Animatics Before a Live-Action Shoot (read 2026-10-13)

SceneMixer turns a novel into a watchable short-drama series in one pipeline: upload the novel (Chinese or English) → AI extracts characters, scenes, and an episode-by-episode structure → generates consistent character and scene art → writes storyboard scripts → renders video segment by segment (16:9 by default, 9:16 in one setting) → composites full episodes.

SceneMixer, Turn a Novel into a Video Series with AI (read 2026-10-13)