---
id: keras
version: "3.15.1"
license: Apache License 2.0
license_treatment: permissive
maintenance: active
---

# keras — Multi-backend Keras

License: permissive · Maintenance: active · Downloads: 19.5M/mo

## What it is and what it does

Keras 3 is a high-level deep learning framework that abstracts away backend differences, letting you write model code once and run it on JAX, TensorFlow, PyTorch, or OpenVINO. It provides a familiar API for building and training neural networks across domains—computer vision, NLP, audio, timeseries, and recommendation systems—while letting you choose the backend that best fits your performance or deployment needs.

The framework is designed as a drop-in replacement for tf.keras when using the TensorFlow backend, and supports both high-level Keras workflows and lower-level custom components. It depends on numpy, h5py, rich, absl-py, namex, optree, ml-dtypes, and packaging. You must install and configure a backend separately before use, and the backend cannot be changed after import.

Use it for:

- Build computer vision models (CNNs, transformers) that run on JAX for research or on TensorFlow for production.
- Migrate existing tf.keras code to run on PyTorch or JAX without rewriting the model definition.
- Train NLP models (text classification, sequence-to-sequence) with a unified API across multiple frameworks.
- Deploy inference-only models using the OpenVINO backend for edge or embedded systems.
- Write backend-agnostic custom layers and metrics that work across TensorFlow, JAX, and PyTorch.
- Prototype on CPU with PyTorch, then scale to TPUs with JAX using the same model code.

## Worth the install?

AI-flagged interpretation of the facts above — verify before relying.

Keras 3 is a multi-backend deep learning framework supporting JAX, TensorFlow, PyTorch, and OpenVINO, enabling you to build and train neural networks for computer vision, NLP, audio, timeseries, and recommender systems.

Yes. Keras 3 is actively maintained, widely adopted (19.5M monthly downloads), has no known vulnerabilities, and offers genuine value if you want to avoid framework lock-in or need to switch backends. Install it if you're building deep learning models and want flexibility; skip it only if you're committed to a single framework and don't need Keras's high-level API.

## Install

pip install keras
uv add keras
poetry add keras

## Installing keras

Before you install: Installation is straightforward with low friction—a pure Python wheel with eight runtime dependencies. The package is actively maintained with a recent release and high community adoption, though you must separately install at least one backend (TensorFlow, JAX, PyTorch, or OpenVINO) for the framework to function.

License in practice: Licensed under Apache License 2.0 (permissive), so you can use, modify, and distribute Keras 3 freely in commercial and open-source projects with minimal restrictions.

Quickstart:

pip install keras

import os
os.environ["KERAS_BACKEND"] = "jax"  # or "tensorflow", "torch", "openvino"

import keras
model = keras.Sequential([keras.layers.Dense(10, activation="relu")])

You must install a backend package (tensorflow>=2.16.1, jax>=0.4.20, torch>=2.1.0, or openvino>=2025.3.0) separately, and set KERAS_BACKEND before importing keras.

Verify before relying:

- Actual performance gains (20% to 350% speedup claims) depend on model architecture and backend choice—not independently verified here.
- Compatibility with custom tf.keras components and conversion effort required for non-trivial models.

## Package facts

- License: Apache License 2.0 (permissive)
- Python support: supports_current
- Install friction: low
- Maintenance: active
- Downloads: 19.5M/month (top 5,000 on PyPI)
- Known vulnerabilities: none known

## Tags

deep learning framework, neural network library, multi-backend machine learning, keras tensorflow pytorch jax, high-level deep learning api, model training framework, computer vision nlp audio, multi-backend, deep-learning, neural-networks

[View on SkillFed](https://skillfed.io/packages/keras) · [View on PyPI](https://pypi.org/project/keras/)