Research · Note · May 2026
The 16 MB fit is vocabulary limited
Not n-gram limited, which changes what you should spend the next megabyte on.
The finding
The constraint is not where people assume
The intuition is that a small artefact is small because it holds fewer n-grams. On a 16 MB build that is not the binding constraint. Pruning at count two or more already removes 55.1% of trigram mass, so the trigram order is nearly vestigial and cutting it further frees little.
What binds is vocabulary: the lexicon, the qualia relations and the fingerprints that index them. That is where the next megabyte should go, and it is where the one after that should go too.
Consequences
What this changes in practice
It means a 16 MB build is honestly described as a bigram model with trigram exceptions, and we describe it that way rather than calling it a trigram model. It also means the way to improve a small build is to author more domain, not to tune the language model.
For a design partner this is good news: authoring effort converts directly into capability at the small end, which is exactly where a regulated embedded deployment lives.
Computed, not measured
These proportions are computed from a corpus analysis rather than measured on a shipped device. The figure that is measured is the trigram hit rate on the held-out set, which is reported separately in the development log.