The most important change in this year's flagship phones is not a camera or a screen. It is that a growing share of their AI features no longer round-trip to a data centre. Summarisation, translation, photo editing, voice dictation and message drafting increasingly run on the phone's neural processor, and the trend is accelerating for reasons that are part engineering, part economics and part regulation.
Small models got good
The enabler is the small language model. Two years ago, models small enough to run comfortably on a phone were toys — fine for autocomplete, useless for anything nuanced. A series of architecture and training advances changed the equation: distillation from larger models, better quantisation, and benchmarks designed for on-device constraints rather than cloud-scale ones. The current generation of sub-five-billion-parameter models handles summarisation, structured extraction and conversational assistance at a quality most users cannot distinguish from cloud services for everyday tasks.
- Translation and dictation now run offline on all major platforms
- Photo AI — search, edit, enhance — is almost entirely local
- Messaging suggestions and summarisation default to on-device
The economics flipped
Cloud inference is expensive for the companies, too. Every query answered by a phone is a query nobody pays a data-centre bill for. As AI features became expected rather than premium, the unit economics of serving them from the cloud at zero marginal price stopped making sense — which is why phone makers spent their silicon budgets on neural engines and shipped the results quietly, inside feature lists rather than keynotes.
What it means for you
Practically, three things. Privacy: local processing means your voice notes and messages need not leave the device, and the settings now exist to enforce that. Battery: neural processors do this work more efficiently per task than radios streaming it to the cloud, though heavy use still costs more than not using AI at all. And capability gaps: the smartest model you can reach is still in a data centre, and phones now fall back to it gracefully — a design pattern, tellingly, borrowed from the offline-first app era.
The longer-term question is where this ends. If the phone's neural engine can hold a useful model, so can the laptop, the car and the earbuds — and the personal-computing stack becomes a federation of small local models rather than one big distant one. That future has real limits, and they are worth remembering: on-device AI is brilliant at tasks about your content and mediocre at tasks about the world. The cloud is not going away. It is just being asked to do less.