Reinforcement learning environments
Open RL environment hubs
Last updated /Reinforcement learning environments
An environment hub is a public shelf of packaged environments: each one described, versioned, and runnable without a private thread. Two shelves people mean are Prime Intellect's Environments Hub and the OpenEnv Hub on Hugging Face. This page says what those projects publish about packaging. It is not a mirror of their catalogs, and a package listed there is not something we sell.
The field guide is RL training environments: a field guide. The pillar is reinforcement learning environments: what they are and what real work adds. The interface those packages often speak is Gymnasium-style APIs for agent environments. How to read a benchmark as a template, which is not the same as browsing a hub, is agent benchmarks that show how environments are built.

What an environment hub is
A hub lets you see the contract before you pay: the version, how to install or run it, and where the reward is computed. A public echo environment proves a client works. It does not prove an agent can do your job. Use a hub to learn the packaging. Use a spec when the task is yours.

Prime Intellect Environments Hub
Prime Intellect describes the Environments Hub as a community platform for environments used in reinforcement learning and evaluation, and as a Python package registry. Their docs say environments are modules with dependencies in a project file, distributed as wheels, and that the hub is moving newer environments onto the verifiers v1 API while older packages are marked legacy. The docs are https://docs.primeintellect.ai/tutorials-environments/environments . The launch post is https://www.primeintellect.ai/blog/environments . Read the version and the runtime on the package you install. A legacy tag is a fact about that package, not a quality score we assigned.
OpenEnv Hub on Hugging Face
Hugging Face's launch post describes the OpenEnv Hub as a shared place to build and explore environments that follow the OpenEnv specification, including a look at the tools and the observations: https://huggingface.co/blog/openenv . The project space is https://huggingface.co/openenv . The reset and step contract those environments use is the API page linked above. A space you can open in a browser is still a package with a version. It is not a custom environment built on your records.
Packaging and versions
Both hubs are about packages you can name and install, not about a single unversioned demo. Pin the version before you compare two runs. A hub that lets the environment compute its own reward is useful only if you can find that function. A listing with no version and no reward code is a brochure. The grid on this page shows version tags as the idea. The tags are not a copy of either catalog.
When a package moves from an older interface to a newer one, treat that move as a new delivery. Re-read the reward and the reset. Do not assume yesterday's install instructions still start the same world. The docs linked above are the place to check that claim, not a screenshot of a gallery. A gallery is not a version you can pin.
Open vs custom environments
An open environment is one you can inspect and run from a public package. A custom environment is one built for a domain, often with someone else's operating history and a contract. The open package will not contain your tickets. The custom package should not pretend to be a public listing. If you need records from a company, say so on the buyer request. A request does not mean a hub package matches that need.
- Read the version and the install instructions before you train.
- Find reset, step, and the reward in the package, not in a blog summary.
- Treat a legacy or runtime label as something to check, not as our endorsement.
- Do not confuse a public echo task with your workflow.
We do not operate either hub, and we do not certify the packages on them.
Questions people ask
Do you mirror these hubs?
No. The links above are their docs and launch posts. This page is not a catalog.
Is a Hugging Face space a custom environment?
Not by being on the hub. Custom means it was built for a named domain and a contract. A public package is the other thing.
What does the legacy tag mean?
Prime Intellect's docs, linked above, use it for environments still on the older verifiers API. Read their page before you rely on a package.
Are the version tags on the image real packages?
No. They show that a hub listing should have a version. They are not names we inventoried.
Can we train on a hub package and skip a spec?
You can run a public package. You still need a spec when the task is your own work.
Will you install a hub environment for us?
No. This page says what to read. It is not a setup service.
Describe the data you need
Tell us the use, the records, and the limits. A request is not a promise that a matching dataset is available.
Tell us what data you need