Hi Product Hunt 👋 I built Onira because the privacy story in this category bothered me. Hypnosis and meditation apps ask you to type the most personal thing in your week into a text field, and then send it to a server. The obvious fix is to not have a server. So Onira runs a 2B open-weight model (Gemma, via LiteRT-LM) on the phone itself. That decision cost more than I expected: the weights are a 2.6 GB download on first run, they sit in the native heap rather than mmapped, and on a mid-range device generating a whole 5,000-word script in one call takes long enough that nobody would wait. The session is generated block by block instead — induction and emergence come from templates, the middle is generated 300-500 words at a time, and narration starts on the first block while the rest is still being written. The model is unloaded the moment the last block lands, because narration then runs for half an hour and that is exactly when you put the phone in your pocket. What I would still like to hear from you: - Does an on-device model actually change whether you would type something real into that field, or is "it's private" a claim nobody believes any more? - The first-run download is 2.6 GB. That is the single biggest thing standing between install and first session. Is there a framing that makes it acceptable, or is it just a wall? Happy to go into the memory side of this in the comments — that turned out to be the binding constraint, not speed.

Onira — Self-Hypnosis

A hypnosis session written for you, offline, by your phone

A hypnosis session written for you, offline, by your phone