Benchmark: Measure What Your Phone Actually Does
Prompt-processing and generation speed, memory headroom and thermal behaviour, per model, on your hardware — and the results feed back into the app's own estimates.
Every claim about local AI speed is made about somebody else's phone. The benchmark screen settles it on yours.
What it measures
Two numbers that matter and are usually conflated. Prompt processing is how fast the model reads what you gave it — the wait before the first word appears, and what dominates when you paste a long document. Token generation is how fast it writes, which is what you watch. A phone can be quick at one and slow at the other, and knowing which tells you whether a longer context is affordable.
Alongside those it records memory headroom actually observed under load, and how the phone's thermal state moved during the run.
Why the app cares about the result
The numbers are not just for you. The selector's estimates of runtime RAM and speed are conservative defaults derived from the model's size and your device tier. Real measurements replace those estimates for your specific hardware, so subsequent model recommendations are based on what your phone did rather than what its RAM figure suggested it might do.
What to do with it
Run it on the model you use, then on the next size up. If the larger one is still comfortably above reading speed, keep it. If it drops the phone into a hot thermal state within a minute, you have learned something a spec comparison could not tell you — and BatteryGuard would have throttled it eventually anyway, more politely than the operating system would have.
Looking for something else? Every page on this site.