pfnstudio

PFN Studio is a studio for prior-fitted foundation models. It is built on the prior-fitted networks architecture introduced in Müller et al., ICLR 2022, and it is aimed at turning a prior into a model that can do in-context inference on real data with the same shape. The page presents it as a single place to design priors, compose architectures, train, test, and share models for a domain.

Its workflow centers on priors, described here as Python code that generates synthetic training data. The studio includes a marketplace of priors and model templates, a cloud runner that trains on the provider’s infrastructure, a public sharing feature that generates Try-it links for completed runs, a beta designer for authoring priors visually, in math, or in Python, and a beta composer for building architectures from blocks. The composer mentions tabular embedders, transformer encoders, attention pools, and task heads, and it also allows custom blocks through a single decorator. The marketplace lists ready-to-train priors, model templates, and OSS PFN projects that can be forked into a workspace.

The page explicitly names ML engineers, data scientists, and domain experts as intended users. It says ML engineers can avoid hand-rolling training loops and evaluation harnesses for each domain, data scientists can get a working in-context model in minutes, and domain experts can start from a prior that matches their domain without PyTorch. It also refers to use cases across tabular data, time series, and causal structure discovery, including forecasting sensor, demand, and traffic series, classifying tabular churn, fraud, and risk data, and discovering causal structure in process and root-cause tasks.

PFN Studio is offered with a Try free option. During early access, CPU is included and GPU is available on demand as projects scale. The page identifies ProfitOps as the company behind it and says the studio used in production deployments was opened up as PFN Studio.