READER-1 AI MODEL LAUNCH:
1B TOKENS, 1 ACQUISITION
AND EMAILS LIMIT HIT.
ALL IN 24 HOURS.
XERJ Reader-1 launched, and the first day did not ease in. In 24 hours the model ran 1B tokens, Model.Surgery joined XERJ, and we hit our daily email limit. That is the kind of day you hope for when you put a new model in front of the world, and then you have to keep up with it. This post is what that day looked like, what the hard cases taught us, and what we are doing next.
The first day, written down
A billion tokens in a day is a lot of real work passing through a model that just went live. Tokens are not a vanity score. They are pages read, clauses checked, questions asked, and answers returned. When that volume arrives in the first 24 hours, it means people did not wait to try Reader-1 on the documents they already have. They put load on it immediately, and the model took it.
That kind of first day matters because it is a signal, not a rehearsal. A launch can look polished in a demo and stay quiet once the door opens. This one did the opposite. Work showed up, and it showed up at a scale you can feel in a single number: 1B tokens. For anyone watching whether a new model has a place in production search, that is the start of a demand curve, written in one day.
1B tokens in 24 hours.
1 acquisition in 24 hours.
All of this in 24 hours.
Our daily email limit was hit.
The same 24 hours held two more facts. Model.Surgery joined XERJ, which we write here as 1 acquisition, and our daily email limit was hit. Those are different kinds of intensity: a research team choosing to build with us, and a mailbox that could not keep up with the day. We will come back to both. The number we want you to remember first is the one the model itself earned: 1B tokens, in 24 hours.
When the page has no answer
Once the tokens started moving, the hard cases were easy to see. They were not the neat English excerpts you put in a launch note. They were questions and documents in languages other than English, and documents where the text has no clear answer at all. A contract written in Spanish. A memo that never names a fee. A filing that talks around a date instead of stating one. That is the work people actually have, and that is where a model for documents has to hold.
Reader-1 answers only when the documents support an answer. If they do not, it replies “Not found”. That sounds like a small product choice. It is the whole point. A text generator is built to keep talking. Ask it for a monthly fee that the page never states, and it will still hand you a number that looks finished. The sentence is fluent. The figure is invented. In a review of a contract, or a search over a records room, that kind of answer is a quiet failure: it looks like a hit, and it is not.
Enterprises, search companies, and teams that build RAG systems need the other habit. They need a model that will stop when the evidence stops. A “Not found” you can trust is more useful than a confident line you cannot check. It keeps a search index honest. It keeps a retrieval pipeline from feeding an invented clause into the next step. It gives a lawyer, an analyst, or an agent a clean gap to go look at, instead of a fake fact that has already started to travel. That is the behavior we built Reader-1 to have, and the first day of traffic pressed on exactly those cases.
The same rule has to hold when the question and the document are not in English. Reader-1 should still read what is on the page, and still refuse to invent what is not. Below is a small interactive example of both situations. It is an example, not a recorded run, and it is there so you can see the difference with your own eyes: one side stops, the other side keeps writing.
Reader-1
A text generator
Model.Surgery joins XERJ
In the same window, the Model.Surgery team decided to join XERJ. They are a research team of AI hackers and mathematicians, including mathematicians with deep topology experience, and they came to build neurosymbolic systems with us: models that propose, programs you can read, and facts that check the answer. That is the same line of work Reader-1 sits on. A model should not be a black box that emits a finished sentence. It should leave a trail you can inspect.
What they add is focus on small, verifiable models for enterprise data. A lot of the world is betting on ever larger systems, hoping a bigger model will get the clause right by instinct. We are building the other way. A smaller model, sized for the job, that writes something you can check, is what a company can actually put next to its contracts, its filings, and its private records. It costs less to run. It is honest when the page is silent. That is the work Model.Surgery wanted to do, and it is the work we want to do.
This is also why the timing matters. Reader-1 had just gone live, and in that same stretch a research team chose to join the company that shipped it. We want to make something that has not existed before, and we want that work to contribute to new science. If you want the longer story of how they came, and what we plan to build together, it is here: Read how Model.Surgery joined XERJ, and why.
We are sorry
Our daily email limit was hit. Some of you did not get a confirmation email, and we are sorry. That is on us, and we wanted you to hear it plainly.
Every one of you will be invited straight to the platform, with no waitlist, and with 1B free tokens instead of 100M. You should not have to wait because our mailbox did.
It shipped anyway
This shipped with a team of one. There is no large bench behind the launch, and no quiet army finishing the last mile. One person held the model, the waitlist, the site, and the day itself. That is a strange way to meet a billion tokens of demand, and it is also the truth of how this got out the door. It is hard to hold all of this alone. It shipped anyway.
If you are used to launches with a full company around them, this may look unlikely. Good. Unlikely is the point. A new way for AI to search for data does not have to wait for a crowd to exist first. It has to work, then it has to ship, then the rest of the story can catch up. We are not done. We are just getting started, and the first day already asked for more.
What's next starts now. The first day proved the work is wanted. The next days are for going further: sharper on the hard cases, building with Model.Surgery, and keeping the door open for the people who already showed up. If you want in, the prompt and the waitlist are below. Copy it, run it, write to us, come build. Let's go.
One prompt, or the waitlist
If you want to try XERJ on your own files, copy the prompt and give it to your AI agent. Then join the Reader-1 waitlist so we can get you in. To learn who builds XERJ and how to check our claims, read About XERJ.
Set up XERJ for me, following https://xerj.org/llms.txt. Install XERJ with the XERJ Query-1 model and start it. Index the folder I point you to (ask before reading anything outside it) and write three facts from that data. Ask XERJ Query-1 one question about them, and show me the answer, the program and the evidence. Then open https://platform.xerj.ai/#request-access so I can join the Reader-1 waitlist.