An open model arrives as somebody else's defaults, shaped for another language, another regulator, another idea of what a model should refuse. We take the best of them apart, measure what they actually do, and rebuild the parts that matter here. Doing that once is a service; being able to keep doing it is the difference between arriving at frontier capability once and staying there.
The first half we can hand over today, on a block of compute locked to you and never oversubscribed. The second is the lab, and most of our effort goes there.
The work runs on European compute. What matters about it is not the badge on the cabinet. It is that the capacity is committed rather than borrowed, that the queue is ours to order, and that the data stays on this continent because the machines are on it.
Some of that is conviction. Most of it is arithmetic. A loop that wants thousands of attempts needs capacity it can saturate on its own terms, and that capacity has to be fit to hold data which is not permitted to leave.
A lab that does not own its schedule is not really running its own research.
Faberon is the part of Sevren that does the research. It decides what is worth trying, writes the code, queues the jobs, reads what comes back, and lets that shape the next attempt. Version 1 is running now, and it does not stop when we go home.
Most of what it produces is failure, which is the correct ratio. A direction is proposed, built, measured against everything we already believe, and severed. Nothing is precious. What clears the bar gets a date and a name on it, and then it goes out.
A lab that runs this way is not faster the way a quick researcher is faster. It is faster the way a factory is faster than a workshop.
Reworking somebody else's release is what the lab can do today. Training frontier models of our own, in this building, is what it is being built to do. The loop is how we expect to cover the distance between those two, and it is the only part of the plan that gets better without being asked.
Running it is the easy half. The hard half is judgment. A loop that grades its own work will fool itself given the chance: it finds the measure that flatters the change, it walks back down a dead end it has already been down, it takes noise for a result at exactly the scale where checking is expensive. Most of what we build is machinery against that. How a direction earns a larger budget, what a finding has to survive before anybody believes it, and what a person is still the better judge of, are open questions here. They are also the work.
We would rather build the thing that finds the next idea than any idea it finds.
A company that wants frontier open models running properly does not want a per-token meter and a rate card. It wants a fixed allocation, a flat rate it can put in a budget, and somebody whose job it is to keep the best open weights in the world fast and current on it. So that is the arrangement: reserved capacity, the hosting managed, the models kept current on it, the bill the same every month.
Nobody burns every cycle they reserve, and the gap between what is booked and what is used is normally just lost. Here it runs the research. The allocation stays guaranteed and untouched, the slack does science, and the work that comes out of it names the companies whose capacity it ran on.
And some of what the loop is pointed at is the serving itself. What a reserved allocation can actually deliver is not a fixed quantity. It is something that keeps being made faster, worked on by the same machinery that does everything else here, in the layers between a model and the metal it answers from. So the allocation bought this year does more work next year, on the same bill. Those gains go out in public with the rest of it.
That tends to be the part they came for. A European company can put this budget into a metered API and end up with a cost line, or put it here and end up with a position: frontier open models on reserved metal it controls, and its name on the research that European openness produced this year instead of next. These are not customers waiting to see how open models turn out. They are the reason the work gets run.
None of it is a bet that open weights will eventually catch up. They have read the benchmarks. The bet they are declining is the other one: a model they cannot inspect, cannot pin to a version, cannot move, and cannot keep if the terms change next quarter.
This is the rare year when the cautious choice and the ambitious one are the same choice.
Sevren publishes on a rhythm rather than a press cycle. Open weights. Papers, with the method and the numbers still attached. And the research databases the loop built on its way there, which tend to be the most useful thing in the box and almost never the thing anybody releases.
It goes out dated, authored, and detailed enough to be argued with.
We would rather be corrected in public than admired quietly.
The first one is close. It is an attention called HEIMR, which the loop turned up while it was pointed at another problem entirely, and it is being written up now. What it does, and what it costs to run, will be in the paper rather than on this page.