Ray

Ray is an open-source AI compute engine with 44,000+ stars on GitHub that scales Python and machine-learning workloads from a single laptop to thousands of machines.

Overview

Ray began at UC Berkeley’s RISELab with a simple promise: write Python code once and let a distributed runtime handle the cluster. It pairs that runtime with a set of AI libraries — Ray Train for distributed training, Ray Tune for hyperparameter search, Ray Serve for scalable model serving, Ray Data for ML data processing and RLlib for reinforcement learning. That combination made Ray the scaling layer under a large share of modern AI: it is used to train and serve large language models and to run production inference and data pipelines at companies of every size. The same script that runs on your laptop scales to a cloud or Kubernetes cluster with autoscaling and fault tolerance built in, which is why Ray shows up everywhere from research labs to the infrastructure of frontier AI labs. It is Apache-licensed, maintained by a large community together with Anyscale, the company founded by Ray’s creators.

Key Features

• Distributed Python runtime – scale functions and actors across a cluster with minimal code changes.
• Ray Train – distributed training for PyTorch, TensorFlow and other frameworks.
• Ray Serve – production model serving, widely used for LLM inference.
• Ray Tune – large-scale hyperparameter search with state-of-the-art algorithms.
• Ray Data and RLlib – scalable data processing and reinforcement learning libraries.
• Cloud and Kubernetes native – autoscaling clusters on any cloud or on-premises.
• Proven at frontier scale – used for training and serving very large models in production.

💳 Pricing

Ray is free and open source under the Apache License 2.0. Anyscale, founded by Ray’s creators, offers a managed Ray platform with paid plans for teams that want hosted clusters and enterprise support.

Who Is It For

ML platform teams, researchers and infrastructure engineers who need to train, tune or serve models beyond a single machine — and want one framework from experiment to production.