The interesting models of 2026 are not the biggest ones. They are the ones small enough to run where the data already is.
What fits in a pocket
| Task | Good fit on device? | Why |
|---|---|---|
| Summarise a note | Yes | Short context, forgiving output |
| Classify an email | Yes | Closed set of labels |
| Redact personal data | Yes | Data never leaves the phone |
| Long research report | No | Needs retrieval and a bigger model |
Latency is a feature
A round trip to a hosted model costs hundreds of milliseconds before the first token. On device, the first token arrives before the user's finger leaves the button. For autocomplete and inline actions, that difference decides whether the feature feels native.
Measure before you ship
Small models fail quietly. Build a fixed set of real examples, run it on every model update, and read the failures by hand. The eval is the product spec.
Member discussion
Become a member of Offset to join the conversation.