The lab for AI that runs on your phone.
Measure the AI in your pocket.
Speed, heat, accuracy and confidence for models running on your iPhone. On-device chats never leave the phone.
Join the beta on TestFlightFree for iPhone. Larger models need more memory; the app shows which ones fit your phone. You'll need Apple's TestFlight app.

A closer look at every word
Your next question is an experiment.
Illustrative confidence colors. Confidence is not correctness.
From the word to the device
Measure everything.
Watch live tokens per second, time to first token, memory and thermal state. Run endurance tests to see throttling and battery use per 1,000 tokens.
Read confidence on every word. Open the token inspector to see what the model almost said instead.
Llama 3.2 3B: 12.3 tokens per second on an iPhone 12 Pro (measured)
One measured run. Speed depends on the model, device and settings.
Token confidence does not establish accuracy.
Put the answer to the test
Know if it's right.
Evaluate models on MMLU and GSM8K subsets, or bring your own question sets. Read accuracy with 95% confidence intervals. Compare mode puts accuracy against speed.
Check calibration: is the model's confidence honest?
Keep the experiment
Reproduce it.
Every reply is a recorded run, with its seed, settings and device. Rerun it and check that it matches.
Export JSON and CSV. Keep the raw record, inspect it yourself and share what you choose.
Every reply. A run record.
Illustrative record fields
- Seed
- Saved
- Settings
- Captured
- Device
- Recorded
- Time to first token
- Recorded
- Tokens / second
- Measured
- Thermal state
- Tracked
On your device
Private by design.
Models run on your iPhone. No account, no analytics, no ads. On-device chats send nothing.
Once a model is on your phone, you can work offline.
The app goes online only when you download a model or choose a remote provider.
Read the privacy policyMore ways to ask
Every engine, side by side.
Explore open GGUF models through a starter catalog and Hugging Face. Ask vision models about photos on your phone. See reasoning models' thinking separately from their answers.
Run Apple's own on-device model on iPhones with Apple Intelligence.

For organizations
Inspect the run. Keep the record.
Evaluate AI on the device, with no account. On-device runs send nothing. Export reproducible run records for your own review.
Built by Peak Summit Labs.
Talk with us about your evaluationTry the beta
The newest features arrive first in the public beta on TestFlight.
Questions? support@peakpocketai.com