PyTorch vs TensorFlow: Which Should Beginners Learn?

Updated October 2026

01At a glance

02Why PyTorch won the beginner question

PyTorch feels like normal Python. Tensors behave like NumPy arrays, the model is a Python class, and the training loop is a for loop you can read line by line. When something breaks, the error points at your line. That makes it a better teaching tool: you see forward pass, loss, backward pass and optimiser step explicitly instead of hiding them in one call.

The bigger reason is the ecosystem. Research groups moved to PyTorch around 2018–2019, and the open-model world followed. An AssemblyAI analysis of Hugging Face and Papers with Code data found that in 2023 about 92% of models on Hugging Face were PyTorch-exclusive, and roughly 70% of new paper implementations in a recent quarter used PyTorch against about 4% for TensorFlow. In December 2025, Hugging Face released the first Transformers v5 candidate with PyTorch as the sole backend and announced it is sunsetting TensorFlow and Flax support. If you want to read a new paper's code or fine-tune an open LLM, you will meet PyTorch.

03Where TensorFlow still makes sense

TensorFlow is not dead; it is a mature production system with a smaller share of new projects. Honest reasons to learn it:

• →On-device ML on Android. TensorFlow Lite was renamed LiteRT in 2024 and is now part of Google AI Edge. It is widely used in Android apps, and TensorFlow 2.20 moved on-device work out of tf.lite into the separate LiteRT project. PyTorch's answer, ExecuTorch, reached 1.0 in October 2025 and is catching up.
• →Existing production systems. Many companies have TensorFlow models in production behind TF Serving. Someone has to maintain and migrate them.
• →ML in the browser. TensorFlow.js remains one of the easiest ways to run a model in a web page.
• →Your syllabus or team uses it. Some university courses and older books still teach TensorFlow/Keras. Learn what you are graded on, then add PyTorch.

04Keras 3 changes the old argument

Keras used to mean “the easy TensorFlow API”. Keras 3 is a rewrite that runs on JAX, TensorFlow or PyTorch (plus OpenVINO for inference), selected with one environment variable. A Keras 3 model can even be used as a PyTorch module. So if you like the model.fit() style, you can have it without committing to TensorFlow. For a complete beginner, Keras is a gentle way to train a first network; for anything beyond that, you will still want to understand a raw PyTorch training loop.

05Where JAX fits

JAX is Google's NumPy-like library with automatic differentiation and XLA compilation, strongest on TPUs and large-scale training. Google DeepMind has used it widely for its research since around 2020, and libraries like Flax and Optax are built on it. It is excellent, but it uses a functional style (no hidden state, explicit random keys) that confuses beginners. Treat JAX as a later, optional skill if you head into research or work at a TPU-heavy company.

06Choose PyTorch if… / Choose TensorFlow if…

Choose PyTorch if…

• →You are learning deep learning for the first time.
• →You want to work with LLMs, Hugging Face models, fine-tuning (LoRA, QLoRA) or research code.
• →You are aiming for ML engineer, AI engineer or research roles at product companies.

Choose TensorFlow if…

• →You are building Android or embedded features and your team standardises on Google AI Edge tools.
• →You are joining or maintaining a team whose models already run on TensorFlow and TF Serving.
• →Your course or exam is taught in TensorFlow; learn it for that, then pick up PyTorch.

07Switching between them

The ideas transfer completely: tensors, autograd, layers, losses, optimisers, data loaders. Once you can train a CNN and fine-tune a transformer in one framework, the other takes days to weeks to read comfortably, mostly spent learning API names. Going from Keras to PyTorch, the main adjustment is writing your own training loop (or using a light wrapper like PyTorch Lightning). Going from PyTorch to TensorFlow, learn Keras 3 first and use lower-level TensorFlow only when you need it. For moving trained models between them, ONNX is the usual bridge.

08A learning path

• →Get classical ML down first with scikit-learn; see machine learning vs deep learning for why.
• →Follow the PyTorch section of the Machine Learning roadmap: tensors, autograd, nn.Module, datasets and data loaders, then a full training loop with validation.
• →Build two projects: an image classifier and a small text model. Our AIML project ideas list has options at each level.
• →Then go into transformers and fine-tuning in the AI Engineer roadmap (PyTorch, NLP & Transformers, Fine-tuning & Alignment sections).

09Frequently asked questions