Model variant, inside one release
A variant is a sibling published alongside the main model: a smaller parameter count, a distilled version, or a build at a different numeric precision. It is a different model and it needs its own figures. As of 2026-09-22.
| Kind | What it changes |
|---|---|
| Smaller parameter count | The count, the memory and sometimes the licence |
| Distilled build | Speed and output characteristics |
| Quantised build | The numeric format and the memory floor |
| Prime or pro build | Usually nothing published in this register |
Inclusion rule. Kinds of variant that appear in the register. A kind whose differences are not documented anywhere is listed as such. Order. From the kind that changes the most columns to the least.
1A distilled model is not the large one running faster
Distillation trains a smaller network to imitate a larger one. The result behaves differently, and a benchmark that does not say which build it ran is comparing two things without naming either.
Publishing distilled and quantised builds beside a full model is also a hardware statement: it says the vendor expected a range of machines, without committing to a figure.
2Variants are where licences sometimes split
In one family in this register the smaller model carries a standard licence and the larger one carries the project's own. A reader who checked the repository badge has cleared one of the two.
So the variant name belongs in the licence cell as much as in the size cell. This register keeps it in both.
3Which columns variants unsettle here
In one family the licence differs between two published sizes, so the variant name belongs in the licence cell as well as the size cell. In two others the memory figure changes with the variant while the output size is held constant.
The register keeps the variant name inside the value wherever the vendor stated one, because a figure without a named release cannot be tested by anybody.
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: Distillation, Quantisation.