Ray iconRay

Open-source AI compute engine for scaling Python: distributed data processing, training, tuning, serving and reinforcement learning

44k+ GitHub Stars
2.60.0 Latest Release (Nov 2026)
PyTorch Foundation Hosted Project (Oct 2025)

Overview

Ray is "the AI compute engine": an open-source framework that takes ordinary Python and spreads it across a cluster. Ray Core gives you distributed tasks and actors, and a set of libraries sits on top: Ray Data for data processing and batch inference, Ray Train for distributed training, Ray Tune for hyperparameter search, Ray Serve for model serving and RLlib for reinforcement learning. It works with PyTorch, TensorFlow, JAX and XGBoost and runs on laptops, VMs or Kubernetes. Ray came out of UC Berkeley's RISELab; its creators founded Anyscale in 2019, and on October 22, 2025 Anyscale contributed Ray to the PyTorch Foundation under the Linux Foundation. It is Apache-2.0 licensed and the latest release is 2.60.0 (November 4, 2026). Also from Anyscale: Anyscale, the managed Ray platform. See also: PyTorch, whose foundation hosts Ray.

The Verdict

Who Should Use Ray?

Best For

  • Teams scaling Python data, training or serving past one machine without rewriting it
  • Large batch inference and multimodal data pipelines on GPUs
  • RL and LLM post-training jobs that mix many workers and actors

Not Ideal For

  • Workloads that fit on one machine (the cluster overhead is not worth it)
  • Experiment tracking or a model registry (pair it with MLflow or W&B)
  • Teams without someone to run and debug a cluster

What's Great

  • One engine for data, training, tuning, serving and RL
  • Scales code with few changes, per G2 reviewers of Anyscale
  • Framework-agnostic and Apache-2.0, now under neutral foundation governance
  • Proven at very large scale: Ray cites 1M+ CPU cores for online serving

Watch Out For

  • Steep learning curve; docs are not beginner-friendly
  • Distributed failures are hard to debug
  • Running your own clusters is real ops work; managed Ray costs money at scale

Pricing

View all features & details

Libraries

  • Ray Core: distributed tasks and actors
  • Ray Data: data processing and batch inference
  • Ray Train: distributed training
  • Ray Tune: hyperparameter tuning
  • Ray Serve: model serving
  • RLlib: reinforcement learning

Ecosystem

  • PyTorch, TensorFlow, JAX, XGBoost
  • Kubernetes and cloud VMs
  • Managed option from Anyscale

Real-World Usage

Named Users

  • Case studies: Instacart, Amazon, Pinterest, Canva
  • Used by OpenAI for training, per Anyscale and 2023 press

Governance

  • PyTorch Foundation hosted project since October 2025
  • Releases every few weeks (2.58 to 2.60 between October and November 2026)

How It Compares

Feature Ray SkyPilot Kubeflow Metaflow
Main job Distributed Python compute engine Find and launch compute across clouds Kubernetes ML platform Python ML workflows
Training and serving Train, Tune, Serve, RLlib built in Runs any job Trainer, Katib Steps on cloud compute
Where it runs Any cluster 20+ clouds, Kubernetes, Slurm Kubernetes AWS, Azure, GCP, Kubernetes
Managed option Anyscale SkyPilot Platform Cloud distributions Outerbounds (Anaconda)

Sources: ray.io · GitHub · Anyscale pricing · skypilot.ai

Is this your tool? Add a badge