PyTorch

PyTorch is an open-source deep learning framework: a Python library for defining neural networks, computing gradients automatically (autograd), and training models on GPUs and other accelerators. It was created at Facebook's AI research lab (now Meta), which first released it in 2016; its reference paper, by Adam Paszke and colleagues, was presented at NeurIPS 2019.

Since 2022 the project no longer belongs to Meta: it is governed by the PyTorch Foundation, hosted by the Linux Foundation. As of July 2026, the latest stable release is PyTorch 2.13, published on July 8 according to the official GitHub repository.

What sets it apart

Its defining trait is eager execution: a model is ordinary Python code that runs line by line, so debugging a neural network feels like debugging any other program. On top of that, autograd records tensor operations and computes gradients dynamically, even through the language's native loops and conditionals. The NeurIPS 2019 paper sums up the philosophy: an imperative style without giving up performance. That combination explains how deeply rooted it is in research, where iterating fast matters as much as running fast.

From PyTorch 2 to today: compilation and deployment

March 2023 brought PyTorch 2.0 and torch.compile, a function that compiles a model to speed it up without changing its code, while staying backward compatible. According to the official announcement, it worked on 93% of a benchmark of 165 open-source models, with average training speedups between 20% and 36%. Compilation also opened the path to today's deployment story: torch.export captures a model as a stable graph, and on top of it sits ExecuTorch, the runtime for mobile and embedded devices that reached version 1.0 in October 2025 and was folded into the framework's core with PyTorch 2.13. For interoperability with other inference engines, the ONNX exporter remains available.

Ecosystem and governance

Around the core sit official libraries such as torchvision (images) and torchaudio (audio), and much of the industry builds on top of it: Hugging Face's Transformers library dropped TensorFlow and JAX in version 5 to focus on PyTorch, according to its own migration guide. On governance, Meta transferred the project to the Linux Foundation in September 2022 — with AMD, AWS, Google Cloud, Microsoft Azure and NVIDIA among the founding members of the PyTorch Foundation — which in May 2025 became an umbrella foundation and took in vLLM (large language model inference) and DeepSpeed (distributed training) as hosted projects. The 2022 press release already counted more than 2,400 contributors and 18,000 user organizations; we cite no more recent share figures, as no current public source supports them.

When to choose it

PyTorch is the default choice if you start from models published by the research community or the Hugging Face ecosystem, if you want to debug in plain Python, or if your final target is a mobile device via ExecuTorch. Alternatives such as JAX offer advantages in certain functional workflows and on TPUs, and TensorFlow remains present in systems already in production; but for starting a deep learning project today, PyTorch is the entry point with the strongest backing in documentation, tooling and community.