Stan
Stan is an open-source probabilistic programming language for Bayesian statistical inference, first released in August 2012 by the Stan Development Team, a group of 52 contributors including Columbia University statistician Andrew Gelman, Bob Carpenter, Daniel Lee, and Ben Goodrich.
Profile
Stan is a programming language and statistical tool that lets users build and run Bayesian models to make probabilistic predictions from data.
Stan is an open-source probabilistic programming language for Bayesian statistical inference, first released in August 2012 by the Stan Development Team, a group of 52 contributors including Columbia University statistician Andrew Gelman, Bob Carpenter, Daniel Lee, and Ben Goodrich. The software is written in C++, licensed under the New BSD License, and named after Stanislaw Ulam, a pioneer of the Monte Carlo method. Stan allows users to define statistical models as imperative programs that compute log probability density functions, then performs Bayesian inference using Hamiltonian Monte Carlo sampling.
The project has no disclosed revenue, funding, or headcount as a formal entity; it operates as a distributed open-source community with no central corporate structure. Its interfaces span Python (PyStan), R (rstan), Julia, and the Unix shell, and it runs on Unix-like systems, Windows, and macOS. The latest stable release, version 2.39.0, was published on May 19, 2026.
Stan is used in production environments by companies such as Smartly.io, which reported running Bayesian revenue models for thousands of accounts nightly, managing approximately one million euros in daily advertising spend. The platform's ecosystem includes forums, documentation, tutorials, case studies, and a developer wiki, all hosted at mc-stan.org. No recent funding rounds, acquisitions, layoffs, or leadership changes were found in the provided research dossier. The project's primary challenges include ensuring model convergence in production, requiring careful prior specification and non-centered parametrization for hierarchical models.
Who buys this
- Academic researchers in statistics, economics, and social sciences who need flexible Bayesian modeling
- Data scientists and machine learning engineers at technology companies building production forecasting systems
- Pharmaceutical and clinical trial analysts performing Bayesian adaptive trial designs
- Quantitative analysts in finance and insurance for risk modeling and pricing
- Government and policy researchers conducting causal inference and survey analysis
Strengths and what to watch
Strengths
- Expressive probabilistic programming language that compiles to efficient C++ code, enabling complex hierarchical models without custom sampling code
- Multi-language interfaces (Python, R, Julia, shell) and cross-platform support, making it accessible across diverse technical workflows
- Active open-source community with extensive documentation, case studies, and a discourse forum providing peer support for model development
Watch for
- No disclosed revenue, funding, or formal organizational structure; the project relies entirely on volunteer contributions and institutional support, creating sustainability risk
- Model fitting can diverge or crash in production without careful prior specification and non-centered parametrization, as documented by Smartly.io's production deployment
- Competition from commercial Bayesian modeling platforms (e.g., PyMC, TensorFlow Probability, Pyro) that offer similar capabilities with dedicated corporate backing
Key Information
- Industry
- Stat Languages
- Founded
- 2012
Frequently Asked Questions
What is Stan and what is it used for?
Stan is an open-source probabilistic programming language for Bayesian statistical inference. It lets users build and run models that make probabilistic predictions from data. It uses Hamiltonian Monte Carlo sampling and compiles to efficient C++ code for complex hierarchical models.
How do I use Stan with Python or R?
Stan provides interfaces for multiple languages. You can use PyStan for Python, rstan for R, and also interfaces for Julia and the Unix shell. This makes it accessible across different technical workflows and platforms including Windows, macOS, and Unix-like systems.
Who typically uses Stan in their work?
Stan is used by academic researchers in statistics and social sciences, data scientists and machine learning engineers building production forecasting systems, pharmaceutical analysts for adaptive trial designs, quantitative analysts in finance, and government researchers doing causal inference and survey analysis.
What are the main challenges of using Stan in production?
The primary challenge is ensuring model convergence, which requires careful prior specification and non-centered parametrization for hierarchical models. Smartly.io documented that model fitting can diverge or crash in production without these precautions, especially when managing thousands of accounts nightly.
Is Stan free to use and who maintains it?
Yes, Stan is free and open-source under the New BSD License. It is maintained by the Stan Development Team, a distributed community of 52 contributors including Columbia University statistician Andrew Gelman. The project has no disclosed revenue or formal corporate structure, relying on volunteer contributions.
How does Stan compare to other Bayesian modeling tools like PyMC?
Stan competes with commercial Bayesian platforms such as PyMC, TensorFlow Probability, and Pyro. Its strengths include an expressive language that compiles to efficient C++ code and multi-language interfaces. However, it lacks dedicated corporate backing, which may pose a sustainability risk compared to commercially supported alternatives.
Sources
- mc-stan.org — Official project description, features, community resources, and developer links
- en.wikipedia.org — Founding details, release date (2012-08-30), latest version (2.39.0 on 2026-05-19), license (New BSD), and naming origin
- statmodeling.stat.columbia.edu — Smartly.io production deployment details: thousands of accounts nightly, ~1 million euros daily ad spend, challenges with model convergence and prior specification
- www.smartly.io — Smartly.io's technical blog post on using PyStan in production with Celery scheduling and pickled models