Keras
Keras 3: high-level deep learning API that runs the same model code on JAX, TensorFlow, PyTorch or OpenVINO, with KerasHub pretrained models
Overview
Keras 3 is "a full rewrite of Keras that enables you to run your Keras workflows on top of either JAX, TensorFlow, PyTorch, or OpenVINO (for inference-only)." The keras.io homepage puts it as "the high-level API for building, training, and shipping models on the backend of your choice," and now also names MLX among those backends. You write a model once with Keras layers and ops, then pick the backend: Keras says JAX "typically delivers the best training and inference performance on GPU, TPU, and CPU," while results vary by model. KerasHub adds Keras 3 implementations of popular architectures with pretrained checkpoints on Kaggle Models, including Gemma 4, Qwen, GPT-OSS and Mistral. Keras 3.15.0 (June 2026) added export to a native PyTorch nn.Module, LiteRT export from the PyTorch backend, sliding-window attention and cuDNN SDPA for causal attention. It is Apache-2.0 licensed and maintained by the Keras team at Google; founder François Chollet left Google in late 2024 but, per Google, keeps contributing and overseeing the roadmap. Also from Google: TensorFlow, JAX, Gemma and Colab.
The Verdict
Who Should Use Keras?
Best For
- Teams that want readable model code they can move between JAX, PyTorch and TensorFlow
- Users of Gemma and other KerasHub models who want quick fine-tuning
- TensorFlow users moving to a supported path for new work
Not Ideal For
- Researchers writing custom low-level training loops in one framework only
- OpenVINO users who need training (that backend is inference only)
- Teams that expect identical performance on every backend
What's Great
- One codebase, several backends: switch between JAX, TensorFlow and PyTorch
- KerasHub pretrained models, with Gemma 4 support since April 2026
- Export paths to PyTorch modules and LiteRT
- Google points new generative AI work to Keras 3 (with JAX and PyTorch)
Watch Out For
- Performance differs by backend and model; Keras notes non-XLA TensorFlow is sometimes faster on GPU
- OpenVINO backend is inference only
- Backend lists differ between keras.io (names MLX) and the Keras 3 docs, so check support first
Pricing
View all features & details
Key Features (3.15)
- export_torch: export a Keras model to a native PyTorch nn.Module
- LiteRT (TF Lite) export with the PyTorch backend
- Sliding-window MultiHeadAttention and GroupedQueryAttention
- Causal attention on cuDNN SDPA
- MultiOptimizer and new ops (unique, pinv, matrix_rank, percentile, ssim)
- 3.15.1: security hardening and Python 3.14 support
KerasHub
- Keras 3 implementations of popular architectures
- Pretrained checkpoints on Kaggle Models
- Featured: Gemma 4, Qwen, GPT-OSS, Mistral
Real-World Usage
Google Backing
- Google: "continued investment in Keras 3 demonstrates our commitment to support major ML frameworks"
- KerasHub ships Google's Gemma 4 with multimodal and agentic guides (April 2026)
Community
- Large GitHub following (see badge) and about 19.8k forks
- Two release lines in 2026: 3.15.x and a 3.12.x maintenance line
How It Compares
| Feature | Keras 3 | PyTorch | JAX | Transformers |
|---|---|---|---|---|
| Level | High-level, multi-backend | Framework | Low-level transforms | Model definitions |
| Backends | JAX, TensorFlow, PyTorch, OpenVINO | Itself | Itself (XLA) | PyTorch only since v5 |
| Pretrained models | KerasHub on Kaggle | Via ecosystem | Via Flax ecosystem | Hugging Face Hub |
| License | Apache-2.0 | BSD-style | Apache-2.0 | Apache-2.0 |
Sources: Keras 3 · Keras 3.15.0 · JAX docs · Transformers v5 blog
Is this your tool? Add a badge