Design priors, compose architectures, train, test and share: one place to build a foundation model for your domain, or for your own operation. No data-generation code to write. No infrastructure to glue together. It is the studio Ceptron is built with.
Built on the prior-fitted network architecture introduced in Müller et al., ICLR 2022. A prior is Python code that generates synthetic training data. Train on it once, and the model does in-context inference on any real dataset of the same shape.
A tour of the whole workflow: design a prior, compose an architecture, train, test and share a model, end to end.
Stop hand-rolling training loops and evaluation harnesses for every domain. Author once, and the studio runs everything.
Get a working in-context model on your problem in minutes: fast enough to be a prototyping tool, not a quarter-long project.
You know the physics or the data better than any ML team. Start from a prior that matches your domain, with no PyTorch required.
The Temporal Causal Foundation Model and Ceptron's process world models were authored as priors, trained and evaluated in PFN Studio. We built the studio because we needed it to build them. We opened it up because the next domain-specific model is yours.
Everything you need to take a foundation model from idea to a trained, shareable artifact, and nothing you don't. Click a tool for a closer look.
Every prior is a Python file plus a parameter spec, validated against a baseline and ready to fork into your project.
y = a·x + b + noiseThe reference PFN prior. Random linear functions with Gaussian noise: the simplest demonstrable PFN training task.
ClassificationTwo-Moons Classificationrandom interlocking half-moons + 0/1 labelsThe classification analog of linear regression. Each task is a fresh draw of the two-moons geometry, so the PFN learns a generic 2D classifier.
Time seriesAR(2) Time Seriesy_t = φ₁·y_{t−1} + φ₂·y_{t−2} + ε_tReal-world-shaped time series. Random stationary AR(2) coefficients per task; the PFN learns to forecast any well-behaved autoregressive series.
CausalLinear SCM Discoveryrecover the graph behind y = Ax + εRandom sparse linear structural causal models. The PFN outputs an adjacency matrix: pure structure discovery, no fitted edges.
Short answers to the questions that come up most. If yours isn't here, talk to us.
A foundation model is any model trained on a lot of data and then reused as the base for many tasks. GPT and Claude are foundation models for language: they learned from internet-scale text and can write, summarize and code. PFN Studio is for foundation models in your domain: tabular data, time series, causal structure. They don't read text; they read your data and do in-context inference (forecast, classify, discover structure) without fine-tuning. Same paradigm, different substrate: GPT trained on the web, while prior-fitted networks train on synthetic priors that you or the community author.
A prior.py file that returns synthetic training samples: random linear functions, random AR(2) time series, random Bayesian outcomes, whatever fits your domain. Fork one from the marketplace or write your own in the Designer.
ProfitOps Inc. builds Ceptron: causal foundation models and process world models for industrial operations such as pulp and paper. Industrial AI is bespoke by default: every plant, product and line needs its own model. To ship at our customers' pace we built a studio for authoring priors, composing architectures, training and sharing trained models. PFN Studio is that internal studio, opened up: the same one our team uses on production deployments.
No. PFN Studio is an independent product from ProfitOps Inc. Prior Labs builds TabPFN, a separate tabular foundation model. We share the same research roots, prior-data fitted networks, introduced by Müller et al. at ICLR 2022, but are separate companies and products. ProfitOps also ships its own models in the PFN family, including the Ceptron Temporal Causal Foundation Model, built and trained in PFN Studio.
No. PFN Studio trains on synthetic priors: Python code that generates data. The trained model then does in-context inference on your real data at run time, without ever being shown a labeled example.
Hugging Face hosts pretrained models built by other people. PFN Studio is where you author the prior, train your own foundation model on it, and deploy it. A different layer of the stack: Hugging Face gives you the finished cake; we give you the oven.
Open the Designer and write your own prior in Python. There are five built-in families to start from (regression, classification, time series, Bayesian, causal), or start blank. The marketplace grows from there: we add new priors regularly and accept community contributions.
Yes. Priors are synthetic, so your real data is never used for training. At inference time your data is sent only to your account's API endpoint, isolated per organization, and is never used to improve the platform.
You can try PFN Studio without a card. For team plans, or to build a process world model of your own operation with us, talk to us.
The training subprocess is portable Python, so pfnstudio run works from any machine. Hosted training is the default and what we recommend; teams with compliance constraints can talk to us about running on their own compute.
No installs. We host the training. Or bring us your operation and we will build its process world model with you.