Verify it yourself.
The November 17 section reports idle and in-meeting chat timings. The October 30 section records a bounded local-chat test of that day’s software 0.9.0 build on the review unit. Earlier October trials and August prototype measurements remain separately dated. These samples do not qualify every shipped unit or establish a response-time promise; your unit’s certificate carries its own results. Made-to-order purchasing is open. current ordering and shipping details are available.
This site makes claims — private, local, honest. Claims are cheap. This page is the numbers behind them and the standard commands that let you check every one of them on your own unit, with no cooperation from us. Nothing below is estimated or projected; where results vary by build, consult the dated test conditions and your unit's own burn-in results.
Local chat after and during a meeting
On an earlier build, after a 60-minute recording ended, three typed questions in an ongoing conversation produced their first visible sentence in 23–30 s and finished in 30–37 s. During the recording, 12 typed questions had a median of 49.6 s to the first sentence and 66.1 s to the complete answer. All 12 ran locally. The questions and conversation lengths differed across turns. Timing covers the chat stream.
A dated sample of local answer times.
The review unit, a Raspberry Pi 5 with 8 GB RAM on a 256 GB microSD, running that day’s software 0.9.0 build and local Gemma 3 4B in Private mode, completed a ten-turn scripted conversation. The warm statistics below cover eight model-generated answers; the first uncached-prompt answer and one deterministic device reply are excluded and identified separately. There were no cloud answers, and every answer was read: all were real answers in the language of the question. This is an exploratory sample, not a customer journey or final shipping-image qualification.
| Measurement | Median | 95th percentile |
|---|---|---|
| First chat-stream text | 19.349 s | 29.039 s |
| Completed answer | 27.732 s | 35.858 s |
The conversation’s first question, whose prompt was not yet cached, took 87 s to first words and 94 s to completion. A separate trial after the model was unloaded from memory (not a restart) took 127 s to first words and 131 s to completion. Each is a single observation. Timing starts with the request and ends at the first nonempty text received from the chat stream; screen rendering time was not measured. The warm percentiles use nearest rank; with eight samples the 95th percentile is the slowest observed answer. Prompt length, context, model residency and competing work can change the wait. These observations establish neither an improvement over a prior build nor a guarantee for your device.
Download the measurement summary and release provenance. It contains exact timing aggregates and hashes of the retained private receipts, without conversation text or owner records. Release identity was read from the device API before and after each run, and the benchmark header records the same build. The November sample above used a different conversation and prompts.
The historical test machine
OLLAMA_NO_CLOUD=1)Measured performance
Method noted per number. The Pi is allowed to be a Pi — we publish what it does, not what marketing wishes it did.
get_throttled=0x0 at every sampleReliability, counted
Every boot runs the boot self-test (update completion, ledger, brain, disk, thermal, services, the web app, code integrity and speaker checks) and appends one line to an on-device tally — reliability here is a running count, not an adjective. Speaker audio is skipped when voice is off or no speaker is connected, and the self-test says so; a skip does not prove the speaker works.
The wire
The central claim: in Private mode, with no messaging or accounts connected, the assistant sends nothing you say out of the box. What does leave, and when, is listed under the exceptions below. Check it with standard tools:
- See every open port:
ss -tulpn— expect sshd 22, the app on 8843 (HTTPS) and 8800 (screen/redirect), mDNS, and inference bound to 127.0.0.1 only. Scan it from another machine; it should agree. - Watch the wire while you talk to it: run
tcpdumpfiltered to non-LAN traffic, then have a whole conversation. Expected, in Private mode with Telegram, email and calendar not connected: no conversation content sent to an external service by the assistant. You will still see clock sync (about once a minute), DHCP and mDNS, none of which carry content. - The kill test: unplug your router entirely. Chat still works. Memory still works. Reboot with no internet — it comes back and keeps working.
- Inspect the workers:
ss -ltnpshows the model server bound to 127.0.0.1; the shipped Ollama service disables cloud features withOLLAMA_NO_CLOUD=1, andsudo nft list rulesetshows the rule that drops every non-loopback packet from theollamauser. The model files sit on disk in the standard Ollama layout.
The exact tcpdump filter (hides only traffic that stays inside your network, so anything leaving the house shows):
sudo tcpdump -i any -nn '(ip or ip6) and not ((src net 10.0.0.0/8 or src net 172.16.0.0/12 or src net 192.168.0.0/16 or src net 127.0.0.0/8 or src net fe80::/10 or src net fc00::/7) and (dst net 10.0.0.0/8 or dst net 172.16.0.0/12 or dst net 192.168.0.0/16 or dst net 127.0.0.0/8 or dst net 224.0.0.0/4 or dst host 255.255.255.255 or dst net fe80::/10 or dst net fc00::/7 or dst net ff00::/8))'
What passing looks like during a full conversation — the kinds of lines you should expect (plus the time server's DNS lookups), none carrying content, none to any PHNTM server (there are none):
IP x.x.x.x.123 > <time server>.123: NTPv4 — time sync (the server set by the installed time-sync configuration)
IP 0.0.0.0.68 > 255.255.255.255.67: BOOTP/DHCP — your router's lease renewal
IP x.x.x.x.5353 > 224.0.0.251.5353: mDNS — LAN name announcement
…repeating as clock sync runs (about once a minute); none of it carries conversation content.
Cloud Boosted answers — off by default and using your own API key — sends your question and the context pulled in for it (memories, document excerpts, switched-on notes, older-conversation excerpts, relevant device readings, the time and time zone, offline-library text, your name and language, and the recent turns) to Anthropic, labeled per-answer in the UI. Run tcpdump during a Boosted answer and you'll see TLS to Anthropic. Legacy Autonomous, available only for a retained Telegram connection or previously saved mode and using your own key, accepts actions through that existing connection: what you ask goes to Anthropic, it may send retrieved email headers and excerpts and calendar results to answer you, and the email and calendar actions you confirm reach the accounts you connected.
A retained legacy Telegram connection carries your messages and the replies through Telegram in every mode, Private included,
and the device keeps a connection to api.telegram.org to receive them. A connected mailbox is checked for new
mail only if you switch on Watch my inbox (off by default); then the device downloads new messages in any mode and
important mail can be announced aloud and sent through Telegram alerts you enabled. Switch back to Private, disconnect
Telegram and the accounts, and conversation content stops leaving the box (clock sync, DHCP and mDNS continue).
Delete the key and it's gone from active settings.
In Boosted (and in Autonomous screen or phone chat), Claude can also search the web through Anthropic with your key: the search words go to Anthropic’s search, and each search is listed in Watch the Wire. Private never searches the web. If you choose Your computer as the Boosted answer provider, your question and its context go to that computer on your LAN instead.
Updates, without dependence
Signed updates come from updates.phntmcore.com once a release is published for your unit’s channel. Daily checks are on by default and can be turned off in Device → Care → Updates; checks send no device identifier or conversation content, though the host sees your IP address and request time. System installs need your device password unless you turn on night installation. You can also send the signed release file from your paired phone or computer, or copy it to a USB stick. Either way the device checks the signature and every file's hash before installing, runs the boot self-test, and attempts a verified rollback if that self-test fails (if the rollback cannot be verified, the device reports that recovery is needed). If the update service cannot be reached, the device keeps working; it never needs our servers to run.
The historical benchmark numbers were measured on the August 2026 prototype. The separately dated October 30 section reports an exploratory latency test; they do not replace the older hardware or reliability measurements. Your unit ships with its own burn-in results. If you measure something different, email the builder — that's a bug report, and it's welcome: [email protected].
Own your AI. Keep your words.
Order your PHNTM One. Each unit must complete an automated burn-in on its shipping build before shipment. $799 once, with no required subscription for local use.
$799 one-time · ships within 6–8 weeks after payment (EU: pre-order) · free US shipping · ships worldwide · 30-day returns