Generation gets the attention, but the other half of video AI — understanding it — is having its own efficiency revolution. Models now reason about hour-long footage, and object tracking that once needed a server runs in real time on a phone.
Two ends of the spectrum
At the large end, vision-language models can now watch and reason about very long videos — summarizing an hour of footage, answering questions about it, finding the moment something happens. At the small end, compact models track and segment objects at video frame rates directly on-device, no cloud round trip.
The interesting move isn't a bigger video model. It's the same understanding, delivered under a real-time, on-device budget.
Where the gains come from
As with small language models, the progress comes from taking deployment constraints seriously rather than shrinking a server model after the fact. Rethink what a video model actually needs to compute, design for the phone's budget from the start, and a surprisingly compact model can match specialist systems across understanding, dense prediction, and vision-language tasks.
Why it matters
On-device video understanding unlocks applications that latency and privacy previously ruled out: live camera features that work offline, real-time analysis that never uploads your footage, interactive tools that respond instantly. It's the same story playing out across AI — capability is escaping the data center and arriving on the devices people actually hold. Video, the most demanding modality, is the strongest sign of how far that shift has come.
