A new shape of foundation model

Build a brain for…

Brains aren’t research artefacts — they’re the high-volume decisions your team already makes, automated. The pattern is the same every time: a prior captures the structure of the decision, and at predict time you pass a handful of your team’s labelled examples as context. Same brain, every team — only the context differs.

💼 Finance / CFO

Invoice matching · match invoice ↔ PO ↔ receipt, flag variances by reason

JE anomaly screen · flag unusual journal entries during close

Cash-flow forecast · daily/weekly position with confidence bands

Collections aging · when will this AR get paid?

and 6 more →

🏭 Operations / Plant

Throughput forecast · next-shift output prediction

Quality drift detector · SPC excursion early warning

Predictive maintenance · time-to-failure from sensor signals

Setup-time predictor · realistic changeover estimates

and 2 more →

📈 Sales / RevOps

Deal-stage forecaster · close-probability this quarter

Pipeline coverage · quarter-end ARR from current funnel

Churn early-warning · accounts matching past churners

Lead-routing classifier · auto-assign by territory + history

and 1 more →

👥 HR / People

Time-to-fill estimator · days-to-hire with confidence

Offer-acceptance scorer · likely-to-accept from terms + candidate signals

Attrition early-warning · employees matching past leavers

Comp-band fairness audit · outliers vs internal cohort

🤖 ML / Engineering

HPO surrogate (BO) · pick the next hyperparameter combo

Learning-curve extrapolator · cancel hopeless runs early

Cost-of-experiment forecast · predict the GPU bill before kickoff

Incident-cause classifier · tag root causes from telemetry

and 1 more →

🧭 Your domain?

Got a high-volume decision your team makes from structured data, with some labelled history, where policy drifts between teams?

Probably PFN-shaped.

See the full menu + the 5-point fit checklist: Brain ideas →

None of these are products

Each row is a shape-of-problem sketch you can take into the wizard, swap your own features for, and train. The brain page lets you edit the prior, the model spec, and the eval. Everything in the marketplace is forkable.

A different way to build foundation models

Trained on what you believe — not what you scraped

Write a prior that captures what your domain looks like. Sample synthetic tasks from it. Train a transformer to do posterior inference over those tasks. No giant corpus. No labelled data. No fine-tuning your brain on the customer’s day-one CSV.

In-context inference, in one forward pass

At inference, the brain reads a handful of example rows alongside your query and answers immediately. No SGD on your data, no gradient steps, no warm-up. Change the context — the brain adapts instantly.

Bayesian by construction, calibrated by design

PFNs approximate Bayesian inference over the prior they were trained on. That means confidence bands are real, not heuristic — they come from the posterior, not from temperature scaling.

Reproducible to the published paper

Every paper-backed prior in the marketplace carries the paper’s training schedule, eval suite, and reported numbers. Pick one, hit Try the reproduction, and the brain page shows ✅ or ❌ against the published claim.

Built on the research

Standing on the shoulders of

PFN Studio implements the prior-fitted-network paradigm introduced by Müller et al., Transformers Can Do Bayesian Inference (ICLR 2022). Every paper-backed prior in the marketplace cites its source and reproduces its headline number.

Bayesian regression

PFNs-reference — Müller 2022. Predict a number with calibrated confidence.

Time-series forecasting

TabPFN-TS — Hoo 2024. Forecast a series given its past.

Bayesian optimisation

PFNs4BO — Müller 2023. Suggest the next experiment to run.

Learning-curve extrapolation

LC-PFN — Adriaensen 2023. Extrapolate a partial curve to its endpoint.

Freeze-thaw BO

ifbo — Rakotoarison 2024. Joint forecasting + acquisition.

More coming

TabPFN, ForecastPFN, Drift-Resilient TabPFN, Mothernet — see the roadmap.

From idea to trained brain in five minutes

1. Sign up and pick a persona. Walk me through it (🎓 Basic mode) for the guided wizard, or Give me the tools (🛠️ Developer mode) for direct access to priors, models, evals, and runs.

2. Pick what your brain should do. Predict a number, forecast a series, suggest the next experiment, extrapolate a curve, show confidence — the wizard auto-resolves which paper-backed prior covers your choice.

3. Practise. Quick taste runs on CPU in ~30 seconds. When you’re ready for the real thing, switch to Marathon on a Vast.ai GPU.

4. Try it on your data. Drop a CSV. The studio infers the schema. One forward pass returns predictions with calibrated uncertainty.

5. Share or fork. Publish your brain to the marketplace, or fork someone else’s as a starting point. Everything ships together: prior, model, eval, run config.

Find your path

🧠 What should I build? The full menu of brain ideas — by role, by job-to-be-done, with the 5-point checklist for whether your problem is PFN-shaped.

🎓 First time? Start here A five-minute walkthrough from sign-up to a trained brain predicting on your real data.

🧪 What can a brain do? The capability picker — what each one delivers and which paper-backed priors provide it.

📊 See it reproduce papers The live leaderboard: every paper-backed prior, the paper's number, and a ✅/❌ verdict for your run.

🛠️ Want the full toolkit? Projects, priors, models, evals, runs — the surface beneath the wizard.

🗺️ What's coming next? Promptable brains, brains that fit themselves to your data, one model per vertical.

📖 Need a term? Brain, prior, capability, speciality, mixture, paper-pinned — quick definitions.

Why this matters

Most foundation models are oracles you can’t audit. A PFN is the opposite — its prior is a Python file you can read, edit, and reason about. Train on what you believe. Verify against what the paper said. Ship a brain that you can defend.