On-device meeting notes
for teams that can’t
let data leave.
LoNotes records, transcribes, and summarizes meetings entirely on the iPhone — every AI model runs on-device. No cloud, no account, no servers. Your recordings, transcripts, and summaries never leave the phone.
LoNotes records a meeting, transcribes it, and runs a language model over the result — entirely on the iPhone, with no round-trips to a server. The audio, the transcript, the summary, and the searchable knowledge index are all created and stored on the device and never sent anywhere. There is no account and no cloud sync, so when the recording stops, the only thing that has moved is text on your own phone.
Three screens, one device.
Press the record control to start a session, and again to stop. A live transcript scrolls as the room talks, each passage captioned and questioned. A private knowledge index reads the files you attach and answers from them. Every step runs on the iPhone.
Press to begin.
Each meeting starts when you press the burgundy wax stamp. Press it again to stop — or, later, to add another session to the same entry.
“We can’t read your meetings — they never reach us.”
A studio rack, not a flowchart.
Four on-device modules. Audio enters the top, the finished conversation comes out the bottom. Each runs locally on the Apple Neural Engine — none of them have a network jack on the back.
Five on-device models. Pick one, or keep them all on hand.
Every model runs entirely on the device. Apple Intelligence is the default — it uses the system model your iPhone already ships with, so there is nothing to download. The other weights live on the phone and load on demand; the recommendation is set in red ink.
| Model | Params | Download size | Est. tokens / sec* | Notes |
|---|---|---|---|---|
| Apple Intelligence ★ | ~3B on-device | Built-in (0 MB) | — (not published) | Default — built-in (Apple-Intelligence devices) |
| LFM2.5 350M (ANE) | 350M | 810 MB | ~52 (measured) | Best speed/quality balance of the downloads |
| Qwen3.5 0.8B (ANE) | 0.8B | 1.2 GB | ~43–48 (measured) | Higher quality, still fast |
| Qwen2.5 0.5B (Text) | 0.5B | 309 MB | ~40–50 (estimate) | Smallest download, lowest quality |
| Gemma 4 E2B | ~2B (E2B effective) | 5.4 GB | ~30 (estimate) | Highest quality, multimodal-capable |
* Measured on an iPhone 17 Pro where marked; the rest are estimates. The library uses EmbeddingGemma — a separate ~270 MB install — so that questions like the one in Fig. 2 retrieve the right transcript span without re-reading every meeting from scratch.
What the studio will do next.
The studio ships small and grows fast. These are the features on the bench right now — if one of them is the reason you would switch, it is close.
Privacy, answered.
The questions enterprise and privacy-conscious teams ask first — about where meeting data lives, what runs on-device, and who can see it.
Does my meeting data ever leave the iPhone?
Do the AI models run in the cloud?
Does LoNotes work offline?
Can IT, Loming, or anyone else access my recordings?
Is it suitable for confidential or regulated meetings?
Can I export my meetings?
Is it free? What about ads?
Will there be an Android version?
Full detail in the privacy policy.
Bringing LoNotes to a team?
Questions about rolling LoNotes out across a team, a security or privacy review, or anything else — write to us directly. We read every message and reply ourselves.