AI Potluck
Contribute
Back to Gap Map Infrastructure / ML orchestration

Ray

Anyscale
open source / Overall score: 5.0(leading)

Framework for building and serving AI and ML applications across distributed compute, scheduling work over many machines and mixed CPU and GPU resources. Ray Core provides task and actor primitives; the libraries above it - Data, Train, Tune, Serve and RLlib - cover batch inference, model serving, distributed training, hyperparameter tuning and feature engineering. Applications deploy as Ray clusters on virtual machines or Kubernetes, and the project is governed by the PyTorch Foundation.

Anyscale was founded by Ray's original creators.

Openness

5 high confidence
5.0
license
Apache-2.0(OSI)
source
public(all Ray libraries OSS)
commercial-tier
Anyscale managed platform and proprietary Anyscale Runtime on top
core features withheld
no — a full 11,885-entry tree carries one Apache-2.0 LICENSE and no enterprise/ or ee/ path, no license-key machinery and no closed dependency

Ray's entire core is published under Apache-2.0, and a full scan of the repository finds no enterprise, ee or commercial path behind it. Anyscale sells a managed platform and its own faster Anyscale Runtime as separate products beside the open project, and nothing in the open tree depends on them.

  • https://api.github.com/repos/ray-project/ray recorded 2026-08-12

    Repo metadata: language Python, license spdx_id Apache-2.0, 43,497 stars, default branch master, size 779,592 KB.

  • https://api.github.com/repos/ray-project/ray/git/trees/master?recursive=1 recorded 2026-08-12

    Full untruncated recursive tree of the default branch master, 11,885 entries. The only license paths are the root LICENSE, python/LICENSE.txt and three vendored third-party licenses (dacite, pathspec, speedscope). No enterprise/, ee/, commercial/ or proprietary/ path. The only anyscale-named paths are CI and release-test tooling (ci/docker/ray-anyscale-*.wanda.yaml, release/ray_release/anyscale_job_runner.py); no closed package the open tree depends on.

  • https://github.com/ray-project/ray recorded 2026-08-15

    Repository page for ray-project/ray: 'Apache-2.0 license' in the sidebar, 43.5k stars, 7.9k forks, public and unarchived. Establishes the licence and that the source is public.

  • https://pypi.org/pypi/ray/json recorded 2026-09-28

    PyPI metadata for the ray package reports "license":"Apache 2.0" and the summary "Ray provides a simple, universal API for building distributed applications".

  • https://raw.githubusercontent.com/ray-project/ray/master/LICENSE recorded 2026-09-28

    The root LICENSE file is the Apache License text, "Version 2.0, January 2004", followed by notices for adapted third-party code under permissive licenses, for example "https://github.com/openai (MIT License)". Searched the body for GPL, non-commercial and Creative Commons: none.

  • https://raw.githubusercontent.com/ray-project/ray/master/python/setup.py recorded 2026-09-28

    The package build script reads its license from "LICENSE.txt", and the base install requirements are open packages such as "msgpack >= 1.0.0, < 2.0.0" and "protobuf>=3.20.3". Searched the body for anyscale, enterprise, proprietary and license key: none.

  • https://raw.githubusercontent.com/ray-project/ray/master/README.rst recorded 2026-09-28

    The README describes Ray as "a unified framework for scaling AI and Python applications" made of "a core distributed runtime and a set of AI libraries for simplifying ML compute", listing Data, Train, Tune, RLlib and Serve. Searched the body for enterprise, commercial, paid, premium, proprietary and license key: none.

  • https://www.anyscale.com/blog/announcing-anyscale-runtime-powered-by-ray recorded 2026-09-28

    Anyscale's announcement of the Anyscale Runtime, a separate runtime it describes as "fully compliant with Ray APIs" and benchmarks against the open project: "it took 772 seconds on open-source Ray vs. 431 seconds on Anyscale Runtime".

  • https://www.anyscale.com/blog/ray-by-anyscale-joins-pytorch-foundation recorded 2026-09-28

    Anyscale's announcement that "Ray, the open-source distributed compute framework for AI is joining the PyTorch Foundation, part of The Linux Foundation." Its founders write "We started Anyscale in 2019 to develop and commercialize Ray."

  • https://www.anyscale.com/product/open-source/ray recorded 2026-09-28

    Anyscale's Ray page presents its product as managed infrastructure beside a Ray that runs anywhere: "Public cloud, private data centers, bare metal, Kubernetes cluster — Ray runs anywhere. Or choose Anyscale, and leave the infrastructure to us."

Adoption

5 high confidence
5.0

Adoption is measured as PyPI downloads of the ray package in the trailing month, the product's declared distribution channel. Ray is the standard open distributed-compute and model-serving substrate under the PyTorch Foundation.

Capability

5 high confidence
5.0

Ray's placement groups atomically reserve bundles of resources across multiple nodes, and Ray's own scheduler then places tasks and actors inside that reservation, packed for locality or spread for fault tolerance. Ray's documentation names that all-or-nothing guarantee, every resource for a trial available at once, as exactly what distributed training needs.

  • https://docs.ray.io/en/latest/ray-core/scheduling/placement-group.html recorded 2026-09-15

    Ray's placement-group documentation: placement groups 'allow users to atomically reserve groups of resources across multiple nodes, a concept commonly known as gang scheduling', with tasks and actors then scheduled into them PACKed or SPREAD. 'Distributed Training (e.g., Ray Train and Ray Tune) uses the placement group APIs to enable gang scheduling. In these settings, all resources for a trial must be available at the same time.'

Verified 2026-09-15