What is here is today's state: the Android app, the source of the PC side and MISUNAI as a base. All of it is self-built and unfinished — the state is below, unvarnished.
| App | 0.5.73 (Android, arm64) — running, changed daily |
|---|---|
| Identity | MISUNAI is the name of her kind; she gives herself her own name after a full lunar year — from what the two of you lived through |
| Base versions | Heart and head for the phone, head and heart adapters for the PC — trained only on hand-written examples and explicitly released thoughts, checked for read-out before upload (downloads below) |
| Backup | Once a day, encrypted on the PC, stored on mi-su.me — restorable only with your own recovery code |
| Statements | When you state something instead of asking, your sentence becomes the answer — the head only writes the question. Personal things stay as they are; health and safety only go into training after your yes |
| Training | A checkpoint every 50 steps; if the PC goes off mid-run, the run continues from there instead of starting over. The small models let the graphics card rest briefly after every compute step — sustained full load coincided with crashes |
| PC status | In the trainer: graphics card, video memory, processor, RAM and what is running. Before every run the PC clears the card and does not start next to a program that loads it heavily |
| Versions | Each of the four AIs shows its version in the app, when it was last trained and how much of what it learned is already in the model; releases run with a step-by-step status |
| Model on the phone | Gemma 3 1B, further-trained here, 812 MB, answers offline |
| PC side | Leitstand, EVO, voice service — Python, runs on Windows |
| Large heart | Gemma 4 E4B with its own trained adapter — for long, personal things |
| Head | Gemma 4 E4B with its own adapter — the same base model as the large heart, for factual questions; it gets nothing personal |
⚠️ This is not a finished product. No Play Store entry, no update mechanism, no support. What is here is the state of an ongoing build.
⚠️ The app is signed with the debug key. That key is publicly known. Two consequences: anyone can build an "update" that Android accepts as the same app — and if this is ever signed properly, an existing installation cannot update; it has to be uninstalled. That loses MISUNAI's memory.
⚠️ Our own models are not included — the base versions are. A further-trained model has partly memorised sentences from its person's diary (measured: 0.89 to 0.93 token accuracy). The base versions are separate runs: trained only on the hand-written examples — where the builder appears neutrally as "Midnight Sun" — and on thoughts explicitly released for the base (today: none). Before training the material is checked against every unreleased thought; after training every model gets 20 trap questions, twice each: 0 hits for all four.
⚠️ To be honest about it: the same questions also found nothing in our own phone model, but did find the builder's first name in our own PC heart — the test can find things, but on its own it does not prove everything. The real safeguard is that nothing private was ever in the training material.
Licences: the base versions are derivatives of Gemma
(Google) and are subject to the
Gemma Terms of Use. The
language anchors come from alpaca-gpt4 (English original
CC BY-NC 4.0) and its German translation — so for private, non-commercial
use only.
Copy the APK to the phone and open it. Android will ask once about "installing from unknown sources" — that is expected, there is no store entry.
On first start the app looks for a .gguf file in its model
folder and says so if there is none:
/sdcard/Android/data/me.misu.midnight_sunai/files/Put a Gemma 3 1B model there as GGUF (q5_k_m, about 800 MB). The file with the newest modification date is loaded — not the one with the highest number. Then restart the app.
Without a PC MISUNAI still answers — just without factual questions, without dictation and without the large heart. With one:
unzip to D:\ki
powershell -ExecutionPolicy Bypass -File D:\ki\INSTALLIEREN.ps1The script sets up the environment and starts the Leitstand on port 8080.
On first start an access token is generated into
leitstand\token.txt.
In the trainer (graduation cap, top right) enter address and token:
http://192.168.1.100:8080 (your PC on the local network)Away from home you need a VPN into your own network. Port
forwarding is the wrong way here — the Leitstand can start training
runs and read the diary, and behind it sits a single token. Tailscale is what
has been tested; FERNZUGRIFF-EINRICHTEN.ps1 adds firewall rules
narrowed to your own network and 100.64.0.0/10.
python daten\bauen.py
python daten\mischen.py
soup train --config soup.yaml --yes⚠️ Do not forget --yes — otherwise
soup asks for confirmation, gets none, and ends immediately with
"Aborted.". The log then looks like a crash.
⚠️ Free up video memory before the run. A loaded head takes about 10 GB; the training will not fit beside it.
Download the base versions and basis.json from below into one
folder, then:
python werkzeuge\basis_einsetzen.py C:\Downloads\misunai-basisIt checks every file against the SHA-256 in basis.json, puts
heart and head for the phone into ausgabe\ (the Leitstand offers
them to the phone) and the two adapters into modelle\gguf\. It
overwrites nothing. The adapters need Gemma 4 E4B:
python werkzeuge\kopf_holen.py. Without a PC: put the phone heart
straight into the app's model folder (section 1).
She then does not start empty but with MISUNAI's nature — and becomes yours from there. She gives herself her own name after a lunar year with you.
The download becomes available once you are signed in and your access is approved. It is the same account as everywhere on mi-su.me.