PyTorch is an open source framework, used from Python, to build, train and test neural networks. It works with tensors, which are tables of numbers, and can run the math on a GPU. It started at Meta and is now run by the PyTorch Foundation. It is not an app or a model site.
A question from the app. Pick an answer.
A job post asks for "experience with PyTorch". What is PyTorch?
That is more like Hugging Face. PyTorch is the code you use to build and train the model.
That sounds like Ollama. PyTorch is not an app: it is a library for programming models.
You use it from Python to build, train and test models. It started at Meta and is now run by the PyTorch Foundation.
Inside, a neural network is a huge number of multiplications. PyTorch gives you the pieces to build it (layers, functions) and does the heavy part: it works out how to adjust each weight so the network makes fewer mistakes. That is training; afterwards you use the trained network to answer.
Data travels in tensors: tables of numbers with several dimensions, like a photo (height, width and colors). With one line you can move those tensors to a GPU, which does many calculations in parallel. It is not required: a small model trains on a laptop CPU.
It was created by Meta's AI research team (then Facebook). In 2022 it moved to the PyTorch Foundation, part of the Linux Foundation, so it does not depend on a single company. It is widely used in research and industry, for deep learning on text, images and audio. Many models on Hugging Face are made for PyTorch.
It is not the only one: JAX and TensorFlow, both from Google, do similar work. Each team picks based on its history, its hardware and its tools; none is "the right one".
An ML Engineer post asking for "experience with PyTorch" wants someone who can train, adjust (fine-tuning) and test models in code, not just use them from a chat or an API. If your job is building apps that call a ready-made model, you may never touch it.