PFN Studio

PFN Studio is a studio for prior-fitted foundation models. It is built around authoring priors, composing architectures, training, testing, and sharing models for a specific domain, with a stated goal of avoiding data-generation code and infrastructure setup.

Its workflow centers on a prior that is described as Python code generating synthetic training data. The studio includes a marketplace of priors and model templates, a cloud runner for training with live loss-curve updates in the browser, and a sharing feature that can generate public try-it links for completed runs. It also includes a designer for authoring priors visually, in mathematical notation, or in raw Python, plus a composer for assembling model architectures from blocks such as tabular embedders, transformer encoders, attention pools, and task heads. The page also names support for forecasting time series such as sensor, demand, and traffic data; tabular classification for tasks such as churn, fraud, and risk; and causal discovery for process and root-cause work.

The product is presented for ML engineers, data scientists, and domain experts. It says ML engineers can avoid hand-rolling training loops and evaluation harnesses, data scientists can get a working in-context model quickly, and domain experts can start from a prior that matches their domain without requiring PyTorch. It also describes prior-fitted networks as training on synthetic priors and doing in-context inference on real datasets of the same shape, without fine-tuning.

PFN Studio is delivered as an online product with sign-in and a free trial. Training runs on the provider’s infrastructure, with CPU included during early access and GPU available on demand as users scale. The page identifies the system as being based on the prior-fitted networks architecture introduced in Müller et al., ICLR 2022, and says it was built by ProfitOps, an industrial AI company. It also notes that the same studio is used internally by the team on production deployments.