Safety classification models that run on 8 GB
1 models with published weights that fit in 8 GB — a phone, a base iPad, an Air. Counted at four-bit quantisation, weights only, leaving the machine about a third of its memory and a gigabyte for context: room for roughly 7 billion parameters. A mixture of experts is counted in full, because it is held in full even though only a few experts compute.
The memory of a phone, a base iPad or an entry-level laptop. What fits is small: models of a few billion parameters, quick and cheap to run, good at summarising, classifying and simple extraction, and out of their depth on long reasoning. This is also where on-device makes the most sense, because the alternative is a network round trip for something that takes a moment.
Models that classify content rather than produce it: what is unsafe, what breaks a policy, what should not be shown. They sit in front of or behind a bigger model, are small and cheap by design, and are the part of a system nobody notices until it is wrong. The thing to check is what the model was trained to catch, because policies differ and a guard tuned for one product's rules will be strict in the wrong places for yours.