Ray Project
Ray is an open-source AI Compute Engine and Python-native distributed computing framework that orchestrates AI, machine-learning, and other distributed workloads across CPUs and heterogeneous accelerators. It is designed to let developers scale from a laptop to thousands of GPUs and to unify data processing, training, tuning, serving, and reinforcement learning.
At a Glance
- AI and machine-learning teams
- Foundation-model developers
- Enterprises building production AI platforms
- Cloud and infrastructure engineering teams
- +2 more
AI Tools by Ray Project
(2)
KubeRay
Kubernetes Operator for Ray
Ray
Open Source AI Compute Engine
Discussions
No discussions yet
Be the first to start a discussion about Ray Project
Latest News
Ray 2.59.0 released; Ray Data LLM and Ray Serve LLM APIs reached general availability/stable status, and token authentication became the default for local clusters.
Ray 2.58.0 released with Ray Serve LLM KV-cache and token-aware request routing improvements.
Ray 2.57.0 released with DataSourceV2 enabled by default and Hash Shuffle V2 improvements in Ray Data.
Ray 2.56.0 released as part of the continuing 2.56 release line.
Products & Services
The distributed runtime with tasks, actors, and objects for scaling Python applications.
Distributed, multimodal data processing for structured and unstructured data such as images, video, and audio, including batch-inference pipelines.
Distributed model-training APIs for foundation models, time-series models, traditional ML, and frameworks including PyTorch and other user-selected frameworks.
Distributed hyperparameter tuning and experiment-management capabilities built on Ray.
Market Position
Ray positions itself as an open-source, Python-native distributed compute engine that unifies the AI lifecycle across data, training, tuning, inference, serving, and reinforcement learning, while scaling across heterogeneous accelerators and clouds. Its closest alternatives vary by workload and include Apache Spark for data processing, Kubernetes plus bespoke orchestration, Dask for distributed Python, and specialized training/serving stacks; Ray differentiates through one runtime and library ecosystem spanning prototyping through production.
Leadership
Founders
Robert Nishihara
UC Berkeley researcher and PhD student in the RISELab/AMPLab and Berkeley AI Research communities; his biography describes interests in applied mathematics, machine learning, and optimization. He later co-founded Anyscale and served as its CEO before moving to product strategy and customer growth.
Philipp Moritz
UC Berkeley RISELab researcher and PhD student who co-created Ray with Robert Nishihara and Ion Stoica; he later co-founded Anyscale.
Ion Stoica
UC Berkeley computer-science professor and RISELab leader; co-founder and first CEO of Databricks, and co-founder/executive chairman of Anyscale.
Executive Team
Keerti Melkote
Chief Executive Officer
Joined in July 2024. Founder of Aruba Networks (acquired by HPE) and former executive at HPE. Industry veteran with deep scaling experience.
Philipp Moritz
Co-founder and CTO
Directs the technical vision and development of the Anyscale platform.
Board of Directors
Founding Story
Ray originated in UC Berkeley's RISELab as a response to the difficulty of scaling emerging AI applications and distributed Python workloads. Its initial vision was to make distributed computing accessible without requiring developers to become experts in complex infrastructure, allowing applications to move from a laptop to a datacenter with little or no code change; the creators later founded Anyscale in 2019 to commercialize and operate Ray.
Business Model
Revenue Model
Ray is open-source software rather than a paid standalone product. Commercial monetization is associated with Anyscale, the managed Ray platform built by Ray's creators, which offers hosted and bring-your-own-cloud deployments and usage-based compute/platform billing.
Pricing Tiers
Hosted managed Ray with self-service onboarding; the public pricing page shows workload examples launched with $3 or $5 credits.
Hosted or BYOC enterprise arrangements; BYOC supports any cloud or region and on-premises deployment, with enterprise SLAs and 24x7 coverage.
Target Markets
- AI and machine-learning teams
- Foundation-model developers
- Enterprises building production AI platforms
- Cloud and infrastructure engineering teams
- Research organizations and data scientists
- Organizations running distributed workloads on Kubernetes
- Distributed Python applications and parallel computing
- Multimodal data processing
- Foundation-model and traditional ML training
- LLM and batch inference
- Model serving and production deployment
- Hyperparameter tuning and experiment management
- Amazon
- Microsoft
- Ant Group
- Intel