Tez is an open-source, lightweight training library for PyTorch that removes the boilerplate from training deep learning models while staying close to plain PyTorch code. Released under the Apache-2.0 license and created by Abhishek Thakur, it has about 1,200 stars on GitHub and was last updated in 2023.
Every PyTorch project repeats the same scaffolding: training loops, validation passes, metric tracking, checkpointing, device placement and mixed precision. Tez (the name means sharp and fast in Hindi and Urdu) packages those pieces into a simple Trainer abstraction: you define your model and datasets the PyTorch way, hand them to Tez, and get a clean training pipeline that handles CPU, single GPU, multi-GPU and TPU setups. The design philosophy is restraint: keep things as simple as possible, make everything customizable, and never hide PyTorch behind magic, so the library suits both fast prototyping and production training. It also ships utilities covering the repetitive parts of most deep learning projects, from callbacks to scheduling. Tez came out of the Kaggle community, where its author is a well-known Grandmaster, and it reflects workflows tuned by competition experience.
Tez grew out of real competition workflows, where the same training script is re-run dozens of times with small changes and any friction multiplies. Its examples cover common Kaggle-style tasks such as image classification and text modeling, showing the pattern of defining a dataset and model in plain PyTorch and letting the Trainer handle epochs, metrics and saving the best checkpoint automatically.
Tez is completely free and open source under the Apache-2.0 license. There are no paid plans; you install it from PyPI and train on your own hardware.
Tez is for PyTorch practitioners, Kaggle competitors and ML engineers who want competition-grade training ergonomics without adopting a heavy framework or rewriting their models into someone else’s abstractions.