Keras
Keras is an open-source neural network library written in Python, first released in 2015 by François Chollet, then a Google engineer.
Profile
Keras provides a high-level, user-friendly API for building, training, and deploying neural network models in Python.
Keras is an open-source neural network library written in Python, first released in 2015 by François Chollet, then a Google engineer. It was designed as a high-level API for building and training deep learning models, initially running on top of TensorFlow, Theano, and CNTK. In 2019, Google integrated Keras as the official high-level API of TensorFlow, and since TensorFlow 2.0, Keras has been the default interface.
The project is hosted on GitHub under the keras-team/keras repository. Keras does not have a corporate parent or reported revenue; it is a community-maintained project with contributions from Google and other organizations. As of 2025, Keras remains one of the most widely used deep learning frameworks, particularly among researchers and developers who value its simplicity and modularity.
The library supports convolutional and recurrent networks, transformers, and generative models, and can run on CPUs, GPUs, and TPUs. Keras is also available as part of the TensorFlow ecosystem, and a standalone Keras 3.0 release in 2023 added support for JAX and PyTorch backends, broadening its interoperability. The project has no disclosed headcount, funding rounds, or formal organizational structure.
Its governance remains informal, with Chollet as the primary maintainer. Keras faces competition from PyTorch, which has gained significant traction in the research community, and from higher-level frameworks like Hugging Face Transformers. A notable security vulnerability, CVE-2025-1550, was disclosed in early 2025, allowing remote code execution via malicious model files, highlighting risks in the library's model serialization format.
Who buys this
- Machine learning researchers prototyping new architectures
- Software engineers integrating deep learning into production applications
- Data scientists building predictive models for classification, regression, and sequence tasks
- Educators teaching deep learning concepts in academic courses
- Hobbyists and students learning neural network fundamentals
Strengths and what to watch
Strengths
- Simple, consistent API design that lowers the barrier to entry for deep learning, widely adopted in tutorials and academic curricula.
- Multi-backend support in Keras 3.0 (TensorFlow, JAX, PyTorch) allows users to switch frameworks without rewriting code.
- Tight integration with TensorFlow, including access to TensorFlow's deployment tools (TF Serving, TF Lite, TF.js) and production infrastructure.
Watch for
- CVE-2025-1550: a remote code execution vulnerability in Keras model file handling, raising concerns about supply chain security and trust in the model format.
- Governance and maintenance risk: Keras is largely a one-person project (François Chollet) with no formal funding or organizational backing, creating bus-factor risk.
- Competitive pressure from PyTorch, which has surpassed TensorFlow/Keras in research paper adoption and industry preference, potentially eroding Keras's user base.
Key Information
- Industry
- AI Frameworks, Tools & Libraries
- Founded
- 2015
Frequently Asked Questions
What is Keras used for in deep learning?
Keras is an open-source Python library for building, training, and deploying neural network models. It provides a high-level, user-friendly API that simplifies creating complex architectures like convolutional networks, recurrent networks, and transformers, making deep learning accessible to researchers and developers.
What is Keras 3.0 and how does it support multiple backends?
Keras 3.0, released in 2023, is a standalone version that supports TensorFlow, JAX, and PyTorch backends. This allows users to switch between frameworks without rewriting code, broadening interoperability and giving developers flexibility to choose the best backend for their specific deep learning tasks.
Is Keras part of TensorFlow?
Yes, since 2019, Keras has been the official high-level API of TensorFlow. With TensorFlow 2.0 and later, Keras is the default interface for building and training models, providing a simpler way to access TensorFlow's capabilities, including deployment tools like TF Serving and TF Lite.
What is the CVE-2025-1550 vulnerability in Keras?
CVE-2025-1550 is a remote code execution vulnerability discovered in early 2025 in Keras model file handling. It allows attackers to execute arbitrary code by loading malicious model files, raising significant supply chain security concerns and highlighting risks in the library's model serialization format.
How does Keras compare to PyTorch in 2025?
PyTorch has surpassed Keras in research paper adoption and industry preference, creating competitive pressure. While Keras offers simplicity and multi-backend support, PyTorch's flexibility and strong community have eroded Keras's user base, particularly among researchers who value dynamic computation graphs.
Who maintains Keras and what are the governance risks?
Keras is primarily maintained by François Chollet, its original creator, with no formal funding or organizational backing. This one-person governance creates bus-factor risk, meaning the project could suffer if Chollet becomes unavailable, raising concerns about long-term maintenance and community support.
Sources
- github.com — Project repository, community, and release history.
- www.sentinelone.com — CVE-2025-1550 vulnerability details: remote code execution via malicious Keras model files.
- www.tensorflow.org — Keras is the official high-level API of TensorFlow; TensorFlow 2.x uses Keras as its default interface.
- aws.amazon.com — AWS customer case studies page (no specific Keras case studies found, but AWS hosts Keras as a supported framework).