NodeValid

Published video-model weights, licences and hardware notes

Quantisation, and the memory it returns

Quantisation stores the same weights in a smaller numeric format. The parameter count does not move; the bytes per parameter do, which is why memory figures for one model can differ by a factor of two or more. As of 2026-09-22.

Same count, different bytesHalving the bytes per parameter roughly halves the weights in memory and leaves the count untouched, which is why the size and hardware columns cannot be merged.Full precisionWeights stored at thetraining formathalve the bytesHalf precisionThe usual publishedconfigurationhalve them againfp8A build for cards thatcannot hold the restThe count cannot see the formatOne release here states a floor for its model and a lower one for an fp8 build.
Fig. 1 The count is constant across the row; the memory is not.
What changes and what does not when weights are quantised. Recorded 2026-09-22.
PropertyEffect of quantising
Parameter countUnchanged
Weights in memoryRoughly proportional to bytes per parameter
OutputChanged, by an amount that depends on the model
Card requirementsThe format has to be supported in hardware

Inclusion rule. Properties affected by a change of numeric format. A property that does not change is listed, because that is the point of the term. Order. From the property that never changes to the one that decides whether a card qualifies.

1The format has to exist on the card

A 6GB figure that assumes bf16 does not transfer to a card that cannot do bf16. One repository in this register names the card generations and the formats together for exactly this reason.

So a quantised build lowers the memory and narrows the hardware at the same time. Both halves belong in a hardware statement and only one of them is a number.

2Why this is a size term and a hardware term

The operation belongs to the file, which is why it is filed here. The consequence belongs to the run, which is why the hardware column carries the figure.

One release in the register states at least 24GB for a model and at least 12GB for a distilled fp8 configuration of it. Same weights, two floors, and the difference is software.

3Where this shows up in the hardware column

Four families name a numeric format beside a memory figure, and one of those states two floors for one model: a higher figure for the release and a lower one for a lower-precision build of it.

Where no format is named, the register does not assume one. Assuming half precision would put a convention into a cell that the vendor never stated, on rows where a reader has least to check against.

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: Autoencoder, Diffusion transformer.