Technical Notes

The technical side, for owners who want it. None of it is needed for everyday use.

everything here can be inspected over SSH — poke at all of it

Hardware

ComputeRaspberry Pi 5, 8 GB RAM, quad-core Cortex-A76
DisplayWaveshare 29990: 10.1″ 1280×800 IPS touchscreen (DSI). Some early units have an EM101 panel: 10.1″ 1920×1200 IPS, 5-point touch (HDMI).
AudioUSB speaker with the 29990 display, which has no built-in speaker; the EM101 panel’s own speaker on early units
MicrophoneUSB microphone (included): recording when you tap Talk, plus optional local “Hello Phantom” detection (off by default)
CameraBuilt-in camera (an OV5647 module) on units that have one; a USB webcam also works. It starts only when you use it; see Cam.
KeyboardMini wireless keyboard with thumb touchpad (included; Alt+M = talk shortcut; Tab moves focus)
Power27 W power supply (100–240 V, US plug), included. It plugs into the display, which powers the Raspberry Pi 5. Plug adapter by destination; see In the box.
CoolingDepends on the case: the 29990 aluminium case cools without a fan; units with a fan run it from the Pi 5’s own temperature thresholds.
Storage256 GB microSD card, installed
NetworkGigabit Ethernet + Wi-Fi (2.4/5 GHz)
Battery / clockNo battery, UPS or clock coin cell. Time comes from NTP; after a power cut with no internet, the time may be unknown until network time returns or you set it in Settings → Voice, screen and language → Date & time.

Software stack

OSDebian 13 (trixie), Raspberry Pi kernel, systemd
InferenceOllama / llama.cpp, bound to 127.0.0.1 only, Ollama cloud features disabled with OLLAMA_NO_CLOUD=1
Chat modelphntm-one:chat: Gemma 3 4B, built text-only on the device (3.9 billion parameters; the unused vision part is left out, about 850 MB less memory, the same answers), Q4_K_M, 8K context. The full build stays on disk as phntm-one:chat-vision, the copy to go back to.
Fallback modelllama3.2-phntm-one:fallback: Llama 3.2 3B, Q4_K_M (used when memory is short; chosen at start-up from the memory budget)
Embeddingsbge-m3 (local search over memories and documents)
Speech-to-textwhisper.cpp on the device, run for each utterance and then closed: multilingual Whisper base q5_1 for English and Spanish; Italian uses a locally trained large-v3-turbo q4_0, which takes longer per utterance (or the base model if that file is not installed). The separate, optional wake detector listens all the time when it is on (off by default).
Voice outPiper TTS, on the device: English en_US-libritts_r-medium offers Standard (speaker 0), Bright (12), Warm (828) and Deep (498); Italian it_IT-serena-medium and Spanish es_MX-claude-high have one voice each. All three languages offer three speaking speeds and a preview in Settings; setup can play a spoken introduction.
AppPython/FastAPI (uvicorn), one process (python -m phntm_one.serve) on port 8800 (the screen) and 8843 (phones, HTTPS)
FaceFirefox ESR (kiosk window, user unit phntm-one-face) in a minimal desktop session showing http://127.0.0.1:8800/screen. Settings → Advanced… → Desktop → Open desktop opens the desktop; tap Return to PHNTM on the desktop to come back.
AppsSettings → Apps (or the Apps tile on the desktop): Files, Text Editor, LibreOffice, Browser and Terminal open in their own window, as your own account, in front of the assistant. The Browser has Firefox’s Safe Browsing (phishing and malware warnings), tracking protection and certificate revocation checks (OCSP) switched off, so it does not contact Mozilla or Google in the background; it will not warn you about a dangerous site or download. The desktop panel shows which link is up (Ethernet, Wi-Fi, both) and how far it reaches: green only when the internet was confirmed by the clock sync, pale when nobody has checked, amber for local-only or a sign-in page, grey when offline. A sample document to try ships in Documents/Try PHNTM-One.

Where things live

Services

unitrole
phntm-oneThe assistant: AI, memory, search, web UI (ports 8800 and 8843)
phntm-one-loopsHealth checks and self-repair (see Self-repair), plus briefings
phntm-one-warmLoads the AI models into memory at start-up (runs once)
phntm-one-selftestStart-up self-test: update finished, runtime, the memory budget ledger, brain (the AI: a real embedding, and a real generation unless local AI was already proven in the last 6 hours), disk, thermal (temperature), services, web app answering, code integrity, speaker. Audio is skipped when voice is off or no speaker is connected, and the self-test then says so; a skip is not proof that a speaker works. Setup uses a chime.
phntm-one-face (user unit)The touchscreen face
phntm-one-updateApplies an update you installed; the self-test decides, and it attempts a verified rollback if the self-test fails

Other phntm-* units handle start-up recovery, unlocking the encrypted stores, the repair page and the Telegram connection.

Restart the assistant services: sudo /usr/bin/systemctl restart phntm-one.service phntm-one-loops.service phntm-one-telegram.service. Restart the touchscreen separately, from the device user’s session: systemctl --user restart phntm-one-face.

Network posture

Open ports. This is the complete list (sudo ss -tulpn will agree):

Who can reach what: from your home network, the owner uses HTTPS on 8843 with a paired phone’s credential for memories, notes and documents. Teacher and student accounts sign in separately, with limited access to their assigned learners, and need no phone pairing. Public pages such as this manual and the sign-in pages open without pairing. Control routes (power off, factory reset, restart, account, network changes, the mailbox, approving a skill install) are device-screen only and cannot be reached from the network at all. A guard rejects requests addressed by any name but the unit’s own (DNS rebinding). Nothing answers from the internet: no port forwarding, no relay, no cloud service.

Related topics

From the device manual for software 0.9.0 (November 15, 2026). Your device serves the manual for its installed software; where this copy and your device differ, your device is right.