Distillation, and what it costs
Distillation trains a smaller network to reproduce a larger one's outputs. It lowers memory and runtime, and the result is a different model rather than a faster copy of the original. As of 2026-09-22.
| Property | Shared |
|---|---|
| Licence | Usually, within one release |
| Nominal parameter count | Sometimes, when the count is quoted for the family |
| Memory requirement | No |
| Output behaviour | No |
Inclusion rule. Properties a reader might carry from a parent model to a distilled build. Properties that are not shared are listed, because the assumption is the error. Order. From the properties most likely to be shared to the least.
1Why vendors ship them
A distilled build widens the audience without changing the flagship. It is the cheapest way to answer a hardware question without publishing a figure, which is how one release in this register addresses memory.
It also lowers the cost of a hosted service, which is a commercial reason rather than a technical one and is rarely stated in a repository.
2The comparison trap
Quality claims made against a full model do not describe its distilled sibling, and timings taken on the distilled build do not describe the full one. Both are published under the same family name.
A fair comparison names the build. The register records which builds exist and does not rank them, because ranking would need a benchmark and this register runs none.
3How this appears in the register
One family publishes distilled builds beside a full model and states no memory figure at all, so its hardware cell is empty while its page is not silent. The builds are an answer in a form no column here can hold.
Another states a floor for a model and a lower one for a distilled configuration of it. That is the same information in a shape the column accepts, and it is the clearer of the two.
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: Quantisation, Autoencoder.