PROOF
Shown next to the test that measured it.
Two kinds of reader need two kinds of proof. The engineer who will test us needs results with their tests. The leader who will sign needs cost, like for like. Both are here, and both say how they were measured.
WHAT THE FILMS PROVE
Owned. No tokens. No cloud. Private. In your control.
Recorded in one take each on one phone, with the whole small language model inside it and no cloud behind it. Watch them before you read anything else on this page.
- 25 of 25
- everyday tasks passed on the phone’s graphics chip
- 23 of 25
- on the processor alone
- About 6 s
- a step
- About 90 ms
- to understand a request
Measured by us on a phone.
A film, recorded in one take on the phone. Press play to watch; nothing loads until you do.
ASK ABOUT ANY SCREEN
Read a 2,226-character page and summed it up.
15.4 s
A film, recorded in one take on the phone. Press play to watch; nothing loads until you do.
TWO APPS, ONE ASK
Found three top-rated seafood places and emailed the list.
36 s
A film, recorded in one take on the phone. Press play to watch; nothing loads until you do.
SHOP WITH SOURCES
The three lowest prices, with store and price. Adverts left out.
A film, recorded in one take on the phone. Press play to watch; nothing loads until you do.
IT SEES
Read a bill by sight: the shop, the date, the total. Logged it.
33 s
A film, recorded in one take on the phone. Press play to watch; nothing loads until you do.
IT REMEMBERS YOU
Told once about favourite music; played it when asked.
15 s
A film, recorded in one take on the phone. Press play to watch; nothing loads until you do.
TEACH IT ONCE
Shown a dinner routine once; ran it for a new dish.
47 s
A film, recorded in one take on the phone. Press play to watch; nothing loads until you do.
MOVIE NIGHT
Tickets, calendar and email from one request. Stopped at Pay.
2 min 20 s
Films, muted, self-hosted. Press play; nothing loads until you do.
Every send waits for your tap. Every run stops before payment. Sample bills made for the demo.
- THE STOP YOU HOLD
Volume down stops it, anywhere, mid-step. Every send, payment and sign-in waits for your tap.
- NOTHING LEAVES THE PHONE
The model thinks on the phone’s own chip. What it reads stays there, and every step is kept on the phone for you to see.
OPERATING FACTS
What each product can show today.
- KEY4MB
Seven tasks on film, with their times, above. A small language model, on a phone, fast, private and secure.
- KEY4KB
Every quote checked word for word against the page; the answer says when a quote cannot be found; later amendments known. Shown on a demo library of 38 public local-government orders in English and Malayalam, on film.
- KEY4EDU
Eight instruments; a tutor that runs on the Key4 model and is checked rather than trusted.
- KEY4, THE PLATFORM
Fitted to one job and proved on your test. Evaluations appear below with the tests that produced them.
FOR THE PEOPLE WHO WILL TEST US
Results, with their tests.
Every evaluation we publish comes with the test that produced it, so anyone can run the same test. We publish release notes for every model we lift: the base, the job, the test, the result, the date. Where an independent reviewer has probed a model, their findings and the cases that broke it are published with their consent.
- KEY4, REASONING
Evaluation in progress. The result will appear here with its test.
- AGENT TASKS
Evaluation in progress on a public multi-turn, tool-use benchmark. The result will appear here with its test.
- AN OPEN MODEL, MADE EXPERT FOR DOCUMENTATION
The second base lifted with the method. Release note to follow.
Where this page says “in progress”, it means exactly that. Nothing is claimed ahead of its test.
FOR THE PEOPLE WHO WILL SIGN
The same work, the same test, a fraction of the running cost.
A cost case is honest only when it is like for like: the same task, measured on the same test, to the same bar; total cost on both sides, including our fee and your hardware; real volume over a real period; and the method shown so you can redo it with your own numbers. Below is an illustrative case built from public list prices, labelled as such. Real cases replace it as clients publish them.
An illustrative case
ILLUSTRATIVE · PUBLIC LIST PRICES · OCTOBER 2026
A document-reading task: half a million documents a month, about three thousand words in and three hundred words out for each, answered today by a frontier model through its API.
- RENTED, TODAY
At published list prices for a current mid-tier frontier model, the monthly bill for this volume is in the thousands of dollars; at flagship prices, in the low tens of thousands. Batch discounts halve it; volume growth doubles it.
- HELD, AFTER THE FIRST JOB
A single workstation-class graphics card runs a lifted small model for this task. Its power, at commercial tariffs in the United States, India, the Gulf or Europe, costs tens of dollars a month. The card is a one-off purchase in the low thousands. What we charge is agreed before work starts and does not grow with use.
- THE SHAPE OF THE SAVING
The metered bill, which grows with every question, becomes an agreed cost and a power bill. The crossover arrives in months, not years, and the model, the test and the record are yours either way.
Sources: the providers’ own pricing pages, manufacturer specifications and official electricity tariffs, read in October 2026. Volume is an assumption; put your own in the worksheet.
Bring a problem. Bring an idea. Or pick one up.
One job, your bar, your test. Everything handed over, whether or not you continue.
OWNERSHIP OF THE MODEL, THE TEST AND THE RECORD
No lock. Just the key.
Nothing we build ties you to us. You hold the model, the test and the record, and you can run them without us.
Let intelligence grow with trust, and all of us rise with it.