NodeValid

Published video-model weights, licences and hardware notes

SkyReels-V2 publishes 1.3B, 5B and 14B

The repository publishes 1.3B, 5B and 14B variants. It also states peak memory for the smallest and the largest at the same 540P output, which is the cleanest controlled comparison in the register. As of 2026-09-22.

SkyReels-V2 in the size column, as publishedTwo memory figures for one output size on two model sizes is the rare case where the effect of the parameter count can actually be read off a vendor's own pages.Fixed by the published countLeft to the runThe count1.3B, 5B and 14B variantsWhat it belongs toPublished across several model typesLicence on that releaseskywork-licenseMemory for a runAbout 14.7GB peak on the 1.3B at 540POutput size544 x 960 at 540P, 720 x 1280 at 720PThe output held constant, the count varied
Fig. 1 One output size, two model sizes, and the difference the count makes on its own.
SkyReels-V2 in the size column, with the columns that qualify it. Read from huggingface.co/Skywork on 2026-09-22.
ColumnWhat the pages state
Size, in the vendor's own words1.3B, 5B and 14B variants
The detail published beside itPublished across several model types
Licence, from the same pagesskywork-license
Hardware, from the same pagesAbout 14.7GB peak on the 1.3B at 540P

Inclusion rule. The value this page is about, the detail the vendor attached to it, and the two columns that change how it should be read. A column the vendor left empty keeps its row and says so. Order. This column first, then its detail, then the two columns that qualify it.

1Holding the output constant is what makes it readable

About 14.7GB on the 1.3B at 540P against around 51.2GB on the 14B at the same 540P. Resolution and clip length are the same on both sides, so the gap is attributable to the weights.

Most memory figures in the register vary two things at once, which is why they cannot be compared even inside one family. This pair varies one.

2A ratio, not a law

Roughly a tenfold difference in parameters produces roughly three and a half times the memory. That is a plausible shape for a model where activations do not scale with the weights, and it is one measurement on one family.

Carrying the ratio to another release would be exactly the kind of derivation this register avoids. It is recorded as two figures, not as a rule.

3What the middle variant leaves out

The 5B is published and no memory figure is stated for it, so the interpolation a reader wants is unavailable from the vendor. The register does not fill it in.

Nor does the count settle the length question, which this family answers separately with a frame table running from 257 frames for ten seconds up to 1457 for sixty.

4Sources

The count comes from huggingface.co/Skywork, read 2026-09-22, exactly as the vendor wrote it. Everything else this release publishes is on its family entry; why the count sits in its own column is on size. Nearby: Two sizes, two machines, Counted to the unit.