How Plimm's AI works

Last updated September 28, 2026

Plimm downloads two AI models and a small voice detector to your phone once, 437 MB in all, and runs them there. One writes the reflections; the other turns speech into text. There is no Plimm server, and your entries are never sent to Plimm or any AI service. Put the phone in airplane mode and everything still works.

What runs on the phone

The models Plimm downloads during setup
Model	What it does	Size
Qwen2.5-0.5B-Instruct, 4-bit (Q4_K_M), run by llama.cpp	Reads your entries, writes the reflections and conversation replies, and puts insights into words	379 MB
Whisper base.en (q5_1), run by whisper.cpp	Turns a voice entry into text	57 MB
Silero VAD v5.1.2	Finds the parts of a recording that contain speech	1 MB

All three are open models, and llama.cpp and whisper.cpp are open-source engines. After the download, Plimm checks every file against its SHA-256 fingerprint and does not use a file that fails.

What happens when you write

Your entry is saved in a database on the phone. On Android the database is encrypted with SQLCipher; on iPhone it is protected by iOS Data Protection.
Plimm checks what you wrote for signs of crisis against a list of phrases, which needs no model. If it finds one, it shows crisis resources, such as 988 in the US. It never contacts anyone.
The model reads the entry, on the phone, and notes its emotions and topics. It also flags clear signs of crisis the list missed, which shows the same resources. In a conversation, it writes the reply.
Across weeks of entries, Plimm looks for patterns, and the model puts them into plain sentences.

None of these steps needs a connection.

One model at a time

A language model needs a lot of memory, so Plimm keeps only one model loaded at a time. It unloads the speech model before a conversation, and stops a reply in progress before it transcribes a recording. That is why Plimm needs a phone with about 4 GB of RAM, and why it checks the phone before the download. If a phone cannot run the models well, Plimm says so up front instead of sending your writing to a server.

Check it yourself

After the download, turn on airplane mode. Writing, voice transcription, conversations and insights all keep working. Plimm needs a connection only for the first download, for a model update you choose to install, and to subscribe.

What does leave the phone

A few things do, and none of them contains your writing or your recordings. The privacy policy describes each one in full.

Usage events, sent to PostHog: for example, that an entry was saved, or that crisis resources were shown and at what severity, never your words. You can turn them off in Settings.
App-lifecycle events, such as the first open, sent by the Google Firebase SDK.
Crash reports, sent to Firebase Crashlytics if the app crashes.
A settings check, at most once an hour, with Firebase Remote Config, for values such as the free-plan limits.
Purchase receipts, verified by RevenueCat when you subscribe.
The model download itself, from Hugging Face or Google Cloud Storage.

Limits

English only: for writing, for speech and for the replies.
It is a small model, built to listen, reflect and ask questions, not to advise, diagnose or treat. Plimm is not a therapy app; the disclaimer says more.
There is no account and no sync between devices.
Model updates are offered, never forced. A newer model downloads only if you choose to install it.