PyTorch was released publicly in January 2017 by Facebook AI Research (FAIR). The PyTorch team’s own retrospective blog post “PyTorch, a year in…,” published January 19, 2018, opens with the line “Today marks 1 year since PyTorch was released publicly,” fixing the public debut to mid-January 2017.
In 2019 the team documented the framework formally in the paper “PyTorch: An Imperative Style, High-Performance Deep Learning Library” (Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, and co-authors). The paper describes PyTorch’s defining design choice: an imperative, “define-by-run” programming style where the neural network is built as ordinary Python code that runs line by line, rather than first declaring a static computation graph. As the paper puts it, “every aspect of PyTorch is a regular Python program under the full control of its user.”
This design made models easy to write, read, and debug with standard Python tools, which is why PyTorch spread quickly through the research community. Where TensorFlow’s early static-graph model favored production deployment, PyTorch’s flexibility favored experimentation - and over the following years it became the framework most new AI research was built on, including most large language model work.