TL;DR: PyTorch is the default deep learning framework in 2026. It was downloaded 86.2 million times from PyPI in August against 16.7 million for TensorFlow, and 248,163 Hugging Face models carry its tag against 17,342. Since March 2026, Google has limited TensorFlow to security fixes, bug fixes and dependency updates.
On January 26, 2026, Hugging Face shipped Transformers 5.0 with every TensorFlow class removed.
Thirty-nine days later, Google's own TensorFlow team wrote that it would "exclusively focus on" security fixes, bug fixes and dependency updates, and pointed new generative AI work to Keras 3, JAX and PyTorch.
Most comparison posts still describe the two frameworks as evenly matched.
- 1PyTorch shipped five feature releases in 2026 so far (2.10 to 2.14). TensorFlow shipped one, 2.21, in March.
- 2Not one of the 121 systems in the June 2026 MLPerf Training round listed TensorFlow.
- 3TensorFlow still has twice the GitHub stars: 200,229 against PyTorch's 103,151.
- 4PyTorch retired a tool of its own: TorchServe is archived and gets no more security patches.
Is PyTorch More Popular Than TensorFlow in 2026?
Yes, on every measure of current use. TensorFlow leads only on counts that pile up over time, such as GitHub stars and forks. The four sources below were all read on September 21, 2026.
PyPI downloads
PyPI served the torch package 86.2 million times in August 2026, and tensorflow 16.7 million times, on pypistats.org's count with mirrors excluded. In April the ratio was 3.7 to 1. By August it was 5.2 to 1.
The gap is growing from both ends. TensorFlow's monthly downloads fell 25% between April and August, while PyTorch's peaked at 98.6 million in July.
Download counts include CI pipelines and container builds, so they measure installs rather than developers, for both packages.
Hugging Face models
The split on the Hugging Face Hub is wider still. It lists 248,163 models tagged PyTorch and 17,342 tagged TensorFlow, a ratio of about 14 to 1.
The PyTorch figure undercounts. Most new uploads use the framework-neutral safetensors format (1.39 million models carry that tag), and the libraries that load them, including Transformers 5, now run on PyTorch only.
Developer surveys
In 2022, 12.95% of Stack Overflow survey respondents had worked with TensorFlow in the past year, against 8.61% for PyTorch. By the 2024 survey, PyTorch led, 10.6% to 10.1%. Among professional developers it was 9.1% to 8.8%.
The 2025 survey dropped the frameworks and libraries question, so 2024 is the latest reading. Research-paper shares have no fresh reading either.
Figures like "85% of papers use PyTorch" came from framework trend charts that have no live source now: Papers with Code redirects to Hugging Face since mid-2025.
GitHub activity
Stars reward age, and TensorFlow's repository is nine months older. It holds 200,229 stars and 77,041 forks against PyTorch's 103,151 and 30,003.
Activity tells the other story: PyTorch logged 17,614 commits in the last 52 weeks, TensorFlow 14,322.
Even TensorFlow's commit count overstates human work. In our sample of its latest 300 commits, 89 (30%) came from Google's two automation accounts, "A. Unique TensorFlower" and "TensorFlower Gardener".
Many of the rest touch the XLA compiler, which JAX and PyTorch/XLA also use. Our breakdown of each framework’s open-source community counts the human contributors behind the commits.
Is TensorFlow Still Being Developed?
Only for maintenance. Google announced the change alongside TensorFlow 2.21 on March 6, 2026:
"Going forward, we will exclusively focus on: Security and bug fixes... Dependency updates... Community contributions."
TensorFlow team, Google Developers Blog, March 6, 2026
The same commitment covers tf.data, TensorFlow Serving, TFX, TensorBoard and TensorFlow’s other libraries. The post recommends Keras 3, JAX and PyTorch "for new work in Generative AI".
The 2.21 release itself dropped Python 3.9 and the TensorBoard dependency, and added a few int2 and int4 types to tf.lite.
The Transformers change hit TensorFlow users hardest. The v5 migration guide says the removal lets Hugging Face "focus fully on torch". A team that needs TFBertModel or any other TF class has to stay on a 4.x release.
TensorFlow's on-device work moved out rather than stopped. LiteRT now has its own repository and release line, and converts PyTorch and JAX models as well as TensorFlow ones.
Which Is Faster, PyTorch or TensorFlow?
Speed depends more on the compiler than on the framework, and the published numbers compare compilers inside one framework, not the two frameworks against each other.
The best-documented gains come from PyTorch's own compiler, described in the PyTorch 2 paper presented at ASPLOS 2024.
Those speedups compare compiled PyTorch with uncompiled PyTorch, not with TensorFlow. The paper reports that TorchInductor beat six other compilers on the same suite.
TensorFlow's answer is XLA, which also powers JAX and runs PyTorch on Google TPUs through PyTorch/XLA.
The best public signal of what fast training runs on is MLPerf. Hardware vendors submit their quickest systems to it, and each submission names its software stack.
In the MLPerf Training v6.0 results published in June 2026, 108 of 121 systems ran PyTorch or a stack built on it, such as NVIDIA NeMo, Megatron or AMD Primus.
The rest were four recommender or graph runs on HugeCTR or DGL, one tinygrad entry, and eight systems whose files name no framework. None named TensorFlow or JAX.
That count comes from the framework field of each system file, which we tallied on September 21, 2026.
On phones and edge devices, Google says LiteRT delivers 1.4x faster GPU performance than TensorFlow Lite. PyTorch's equivalent, ExecuTorch, reached 1.0 on October 18, 2025.
PyTorch vs TensorFlow Side by Side
Both ecosystems have lost pieces.
PyTorch's serving tool, TorchServe, is archived, and its README warns that "vulnerabilities may not be addressed". vLLM, the PyTorch-based LLM serving engine, joined the PyTorch Foundation in May 2025 alongside DeepSpeed.
| As of September 2026 | PyTorch | TensorFlow |
|---|---|---|
| Latest release | 2.14.0, September 2, 2026 | 2.21.0, March 6, 2026 |
| Feature releases in 2026 | 5 | 1 |
| Development status | Active, a release about every 2 months | Security, bug fix and dependency updates only |
| Governance | PyTorch Foundation, Linux Foundation | |
| Compiler | torch.compile (TorchInductor, Triton) | XLA |
| LLM serving | vLLM (TorchServe archived) | TensorFlow Serving (maintenance) |
| On-device | ExecuTorch 1.x | LiteRT (now a separate project) |
| Hugging Face Transformers | Supported in v5 | Removed in v5; pin 4.x |
| High-level API | Keras 3 runs on it | Keras 3 (default since TF 2.16) |
Release dates are from each project's GitHub releases page. Keras sits in both columns because Keras 3 runs on JAX, TensorFlow or PyTorch.
That makes it the cheapest exit for a codebase written only against Keras APIs: switch the backend setting to PyTorch and keep the model code.
Which Should You Use for a New Project?
PyTorch, unless you train on Google TPUs, where JAX is the stronger choice. The case for TensorFlow in 2026 is an existing system that works, not a new one.
- A TFX or TensorFlow Serving pipeline is already in production
- The models are Keras, so a later backend switch stays cheap
- You pin a Transformers 4.x release and can live without v5 models
- The work involves LLMs, fine-tuning or anything from Hugging Face
- You serve with vLLM or train with DeepSpeed or Ray
- You want new features: 5 releases in 2026 so far
For a TensorFlow codebase, the risk is dependencies, not the code itself. TensorFlow 2.21 ships wheels for Python 3.10 to 3.13; PyTorch 2.14 also covers Python 3.14, released in October 2025.
Each gap like that, and each model that ships PyTorch-only, adds to the cost of staying. The languages AI teams use have not changed, and Python still sits under both frameworks. Only the framework layer has consolidated.
For learners, PyTorch is the safer first framework. The tools behind LLM work, from Transformers 5 to vLLM and DeepSpeed, run on it, which matters for the AI engineering skills employers pay for and for fine-tuning open models.
How Do You Move a TensorFlow Model to PyTorch?
The route depends on how the model was written. There are three common cases, from cheapest to most work.
- Keras 3 models: set
KERAS_BACKENDtotorchbefore importing Keras. Layers, training loops and saved.kerasfiles carry over, as long as the code calls Keras APIs and nottf.*functions directly. - Hugging Face models: on Transformers 4.x,
from_pretrained(..., from_tf=True)loads a TensorFlow checkpoint into the matching PyTorch class. Save the result as a PyTorch checkpoint, then upgrade to v5. - Custom TensorFlow code: rewrite the model in PyTorch and port the weights layer by layer. This is the slow case, since custom
tf.functiongraphs and TFX pipelines have no automatic converter.
In each case, run the old and new model on the same batch of inputs and compare the outputs before switching traffic.
Small numerical drift from different kernels is normal; a changed prediction on a known input means a weight or layer order went wrong.
Serving is a separate decision. A model that moves to PyTorch still needs a new server once TensorFlow Serving is out of the picture: vLLM for language models, or a general inference server for everything else.
Hiring PyTorch and Machine Learning Engineers
Moving a TensorFlow system to PyTorch, or building a new model pipeline, takes engineers who know both.
Second Talent vets PyTorch developers, machine learning engineers and LLM engineers across Asia, including hands-on use of AI coding tools. For the platform side of that work, see how ML infrastructure and MLOps roles split.
Tell us the stack and we send a shortlist of pre-vetted engineers within 24 hours. Start hiring.
FAQ
Is TensorFlow dead in 2026?
No, but it is in maintenance. Google still ships security fixes, bug fixes and releases for new Python versions. It no longer adds major features, and it recommends Keras 3, JAX or PyTorch for new generative AI work.
Can Keras 3 run on PyTorch?
Yes. Keras 3 runs on JAX, TensorFlow or PyTorch, chosen with the KERAS_BACKEND setting. A model written in Keras 3 can move to the PyTorch backend without a rewrite.
Does Hugging Face still support TensorFlow?
Only in Transformers 4.x. Version 5.0, released on January 26, 2026, removed every TensorFlow and Flax class, so TensorFlow users have to stay on a 4.x release.
Where does JAX fit?
JAX is the framework Google itself recommends alongside PyTorch, and it is strongest on TPUs. It had 7,754 commits in the last year, and it releases monthly.
It also has far fewer published models than PyTorch: 12,125 Hugging Face models against 248,163.
