Founding Members
Your name, in the weights.
Put your name — or a message in your words — on the Founding Roll. If you opt in, it goes into the data the next Brain is trained on. From the next release on, the roll's fingerprint will be written into the model files we ship, and your entry gets a signed certificate that names that model back. £10 buys 500 million training tokens.
As far as we can find, this will be the first AI model file to name its own founders and their certificate registry — with each certificate naming the model back.
📜
The Founding Roll
Your chosen name — or a message in your words — on the roll, in roll order, with everyone who joined. The roll's fingerprint will be written into the model files we ship from the next release, and your entry is trained into the next Brain if you opt in.
Two separate opt-ins, both yours to give or keep: shown on the public roll; used in training. Reviewed by a person before it goes on the roll.
🪪
Your certificate
Every approved Founding Roll entry gets a signed certificate — number, your entry's fingerprint, the roll's fingerprint, and the fingerprint of the ternii model release it is bound to. The model files we ship from the next release will record the roll in return. Signed by Middletech, anchored publicly, and checkable by anyone at ternii.com/founders/verify — including against the model file itself.
If you would like it as a token in your own wallet, just ask and we'll mint one — no wallet needed otherwise.
It is a certificate of membership, nothing more — not a currency, not a share, not something we sell or trade. What is anchored publicly is a fingerprint of each roll version, never your name; if you withdraw, your certificate is marked withdrawn in the registry (Terms, 6a).
🚀
Early builds first
Every Android and iPhone build, and every model file, reaches Founding Members before anyone else.
🛠️
Inside the members' area
The technical film, training notes as they land, the requests board, and the ledger of what your money trained. What's inside →
🧠
What you are funding: Brain 2.0
4.3 billion parameters, and quicker than the one you have. Only about one in five of them fires for any one word — that is what mixture-of-experts buys you — so the model can grow while the work per word goes down.
We have built it and measured it: it is quicker than Brain 1.0 on the same phone, in the same app, and it holds four times the conversation in mind. Decode speed does not depend on what the weights have learned, which is why this can be measured before the model is trained.
On an iPhone 17 Pro, in the app with agent mode and tools switched on — the heaviest thing we ask of it — Brain 2.0 writes at 126–140 tokens a second where Brain 1.0 manages 80. On the bench, away from the app, 2.0 reached 147 and 168 tok/s on two runs. Two reasons it pulls ahead: it reads about 17 % fewer bytes for every word it writes, and it can keep its memory of the conversation in a compressed form that Brain 1.0's shape cannot — a design decision taken months ago, paying off. And when the iPhone 18 Pro arrives we want to know whether 200 is within reach.
Both figures are from the same phone in the same session, warm, generating a real answer — the only thing changed between them is the model. Speed figures move with the length of the prompt and the answer, so we quote ranges. The model is trained only far enough to measure how fast it runs; teaching it is what Founding Members pay for. The 4.3 B file is bigger than today's and will cost more memory to run than Brain 1.0's 195 MB on that phone — we will publish that number when we measure it, like every other one here.
📣
Hear first
If we ever open a round, Founding Members hear about it first.
Subject to eligibility and the law.
One-off. Worldwide. Priced in pounds. The amount shown on the button is what you pay. Where local tax applies, training tokens are credited on the amount before tax.
Names and messages are reviewed by a person before they go on the roll or a certificate is issued: nothing sexual, violent or abusive, no profanity, no other people's personal data. Our decision is final. We'll ask you to change it — or, if you would rather not, refund you in full, whatever stage your membership has reached.
The name can be anyone's — a gift for a child, a grandparent, a pet. Pseudonyms welcome. The person paying must be 18 or over.
We count every £10 as 500 million training tokens — a rate set below what a billion tokens has cost us to train, so it covers failed runs, evaluation, data preparation and price rises. We publish the running total of tokens funded.
Why tokens: the models ternii is compared with learned from trillions — BitNet b1.58 2B4T: 4 trillion; SmolLM3: 11 trillion. Every 200 members = 100 billion more training tokens.
Fees are revenue of Middletech Limited. We intend to spend them on training compute and we publish what we trained, but we do not promise a particular model, release date or result. Early-access builds are pre-release software provided as is.
Payments by Stripe, one-off, in GBP from anywhere; VAT included where it applies.
Founding Membership is a paid membership with the benefits above. It is not a share, a loan, an investment or a donation. It carries no ownership, no voting rights and no financial return. Cancel any time until your entry has been reviewed and your certificate issued; after both, your £10 has bought training tokens and is spent. We always refund a duplicate charge, and an entry we decline. Middletech Limited, company no. 16955822. Terms · Privacy.
terniioffline
