New AI Router Keeps Your Phone Cool and Your Questions Answered
Your phone's AI crashes after a few questions — this fix keeps it running.
Running AI on your phone sounds ideal. It's free, it works without internet, and your messages never leave your pocket. But there's a catch nobody advertises: it runs hot. On a flagship Samsung Galaxy S25+, a researcher found that sustained AI generation didn't just slow down — the graphics chip handling the work crashed or quietly froze after a few questions in a row. It happened even when the phone was cool, and it was worst on long answers. These were not high-end datacenter problems. These were ordinary phone problems.
The fix, called HybridInfer, is less a new AI and more a smart dispatcher. It picks between three options for every question: a small AI living on your phone, a medium AI on a nearby server that can look things up, or a powerful cloud AI. To choose, it looks at two things — how warm your phone currently is, and how hard your question seems. The dispatcher was trained ahead of time using reinforcement learning (trial-and-error learning, like training a dog with rewards), balancing answer quality against speed, cost, heat, and a bonus for keeping things on your phone.
Tested on 210 real prompts on real hardware, the dispatcher gave noticeably better answers than two hand-tuned rules, at the lowest cost of any setup that adapts. Always-on-phone matched quality on questions it could handle, but was three to six times slower and simply failed on long ones — so routing wins on speed, reliability, and coverage rather than raw quality. Notably, without that local-keeping bonus, the system's best strategy was to ship everything to the cloud. Privacy had to be explicitly rewarded, not assumed.
This is one researcher, one phone, and one set of AI models, so treat it as a signpost rather than a finished product. But the direction is clear: your phone is becoming a manager that decides what it can handle alone and what to hand off. That decision will quietly shape how fast your AI feels, how much it costs, and how much of your life leaves your device.
- Phone AI isn't just slower when it heats up — on a Samsung Galaxy S25+ it crashed or froze after a few questions in a row.
- A new 'router' sends easy questions to your phone, medium ones to a nearby server, and tough ones to the cloud, based on heat and difficulty.
- Across 210 test prompts it delivered better answers at the lowest cost, and was 3-6x faster than always answering on the phone.
Why It Matters
Soon your phone will decide what to answer alone and what to send away — changing speed, privacy, and cost.