NodeValid

Published video-model weights, licences and hardware notes

Checkpoint, and why it is not a model

A checkpoint is a saved set of weights: one file, one moment in training, one set of behaviours. A model family may publish several, and a figure stated for the family belongs to one of them. As of 2026-09-22.

Family, release and checkpointThree levels of naming sit between a brand and a downloadable file, and a figure quoted at the wrong level is the commonest error in this field.Widest, and least checkableFamilyA brand covering everything the vendor has published under onename.ReleaseA numbered version, with its own licence and its own date.CheckpointOne file, one set of weights, and where every figure belongs.Figures belong to the bottom level
Fig. 1 The level a figure belongs to decides whether it can be checked at all.
What can differ between two checkpoints in one release. Recorded 2026-09-22.
PropertyCan differ
Output resolutionYes, and one release here states two
Clip lengthYes, stated per checkpoint in the same case
Memory neededYes
LicenceYes, in one family in this register

Inclusion rule. Properties a reader might assume are shared across a release. A property that differs in at least one entry here is listed as differing. Order. From the properties most often stated per checkpoint to the least expected.

1Per-checkpoint documentation is the good case

One release in the register describes two checkpoints separately: one resolution and one clip length each, with a shared frame rate. Nothing has to be inferred about which figure belongs to which download.

Several larger releases state one figure for a family that publishes three variants, which leaves the reader to guess and the register to record the ambiguity.

2Why the register keeps the variant name in the cell

A memory figure without a checkpoint is unfalsifiable: anyone who fails to reproduce it can be told they used the wrong file. Attaching the name makes the claim testable.

It also makes the ladder readable. A family with three published sizes and one memory figure has documented one of its three downloads.

3Where the attribution problem shows up here

Six families publish several variants and attach their figures to named releases. Two of them leave one published variant without a figure of its own, which is the gap a reader hits when they want to interpolate.

Nine families publish a single model, so every figure on those pages has one owner. Narrow releases are more reliable on this axis rather than less.

A reading note, not an entry: no vendor value appears on this page. Where the register records this term for a particular release, it is on the size column and on that family's own page. Nearby terms: Model variant, Distillation.