The working thesis

Focus is the moat.

Pragmatic Frontier is a small lab with two moats: focus, which the large labs cannot buy, and independence, open weights on compute no vendor can reprice or retire. We work hybrid. Closed frontier models and open weights run in one harness, and the evals decide which takes each step. The job is to spend test-time compute well on any long-horizon task: the thinking, tool calls and retries a model burns over hours of unattended work.

The moat

Focus is the one input that does not scale with headcount. A frontier lab splits its attention across everything it ships, and every hire adds a meeting. We are a few people with no meetings and the whole day on the experiments. That is the moat we have today.

The second moat is independence, and the first rack buys it. On a rented model a retired version means every process gets retested, and a price move moves the margin. Our open weights sit on our own disk, where nobody can retire them. The same independence is what we sell.

Owned compute also changes who holds the controls: the turn budget, the goal loop that ends when a check says done, and the heartbeat on a timer. Rented, tokens price the budget, the loop stops at a turn cap before the goal is met, and every heartbeat costs real money. Owned, the budget is what the job is worth, the loop runs until done, and the heartbeat runs all night on GPUs that would otherwise sit idle. The runs nobody would have paid for are where the surprises are. Focus is the input, and owned compute stops anyone metering it.

What we build

A domain-specific harness removes every step between an idea and its outcome: a person states the idea, and the harness delivers the outcome.

Each research target, once met, applies to an enterprise workflow and gives the company full sovereignty: model, harness and data stay on machines it controls.

Integration has its own rules and blockers for what plugs in and when, on a timeline pinned to model releases. Open base models ship on the frontier labs' schedule, and we absorb each one while it is still the best base, or our post-training lands on a model already behind.

How we work

Research comes first and has its own budget. Our output is models and hypotheses, and both need unbroken weeks. The budget exists so we never sell that time to pay the bills.

Repetitive work comes last, and workshops for senior management count as repetitive. Anything that repeats moves into the harness, the agent system that runs our processes, and we reach it wherever the work happens: WhatsApp, email, Signal, Telegram and voice.

Nobody manages anyone. A‑players organize themselves, and we aim to be A‑players who work only with A‑players. Our target is a company worth $10 billion with fewer than 20 people.

The constraint

One risk is building this in the old paradigm. Almost everything written about running a company, from the org chart to the quarterly OKR review, exists to coordinate humans. A lab whose processes run in a harness needs about 2% of it. Copy the other 98% and we import the overhead without the people it was built for.

So until there is a round, paperwork and calls stay at the minimum the work needs. An hour in a meeting is an hour off the moat.

Closed models

The other risk is leaning too far on the closed half. A long-horizon task runs unattended for hours: a research run that reads, trains and evaluates with nobody in the loop. Closed models are still ahead on its hardest steps, and an open model under 30B loses the thread if handed the whole job. So the closed half takes those steps today, and its share of each job is a number we track.

Let that share grow and the lab runs on someone else's terms. A vendor can change the model behind the name we call, and a job in its five hundredth call inherits the change mid-run. A long job also ships its whole context to someone else's servers: a cost on our own work, a broken promise on a client's.

So the harness picks the model per step, not per job. It holds the state and checks each step, so no model has to be good for the whole day. Each step goes to the cheapest model that passes its eval, and an open model takes over the day it passes. Every step on a closed model is a step an open model has not yet earned.

What capital buys

Under €1 million	Rented compute and pay for a small team.
Under €5 million	A rack of our own, with the full stack, post-training included, running on it. The second moat starts here.
Under €10 million	Larger post-training runs.

Small models

Most tasks need a model between 1 and 30 billion parameters. A general model spends its parameters on everything; a post-trained model spends them on one job. We bet that a post-trained 7B beats some general 30Bs on its task, and that a 30B beats a 400B.

The arithmetic is on the small model's side. At 16-bit precision a 7B model is 14 GB of weights and fits on one graphics card; a 400B is 800 GB and needs a rack. That is why the rack in the second tier is enough: every model we train fits on it.