What counts as an open-source RL environment?
An open-source RL environment is one whose sandbox, tasks and grader are published under a license that lets others use them. You can download it, run it on your own machines and train an agent in it without paying anyone.
That covers a wide spread of quality. At one end are carefully built environments maintained by research groups and companies. At the other are weekend projects with a handful of tasks and a grader nobody has stress-tested. The label “open source” says nothing about which end you are looking at.
What is an environment hub?
A hub is a shared registry: one place where people publish environments in a common format, so that others can find them, install them and plug them into a training loop with little glue code.
The best-known example is Prime Intellect’s Environments Hub. Prime Intellect’s own launch post, dated 27 Aug 2025, calls it an open community platform: a place where people publish RL environments and evaluation environments and reuse each other’s. The company also publishes open-source RL training code (prime-rl) and sells a hosted training product, and it runs bounties that pay people to contribute environments. The launch post does not publish the bounty amounts.
A hub solves two problems at once. Builders get a common interface, so an environment written once can be used by anyone’s trainer. Users get a catalog, so they do not start from nothing.
What else is open besides environments?
Training an agent needs more than an environment. It needs the loop that runs episodes, collects rewards and updates the model. Several of these tools are open too.
| What it is | Example from our research | Open part | Commercial part |
|---|---|---|---|
| Environment hub | Prime Intellect Environments Hub | Shared environments, prime-rl | Hosted training product |
| RL training library | OpenPipe's ART library | The library | OpenPipe's tools, now part of CoreWeave |
| Environment engine | Bespoke Labs | Not stated in our sources | Built environments and infrastructure |
Companies named in this table may pay us a fee if an introduction we make leads to work with them. That does not decide who is listed.
OpenPipe makes software that uses reinforcement learning to train agents and publishes the open-source ART library. CoreWeave announced on 3 Sept 2025 that it would acquire OpenPipe. Bespoke Labs builds environments and the infrastructure to run them, and announced $40M of funding on 6 Jul 2026; we did not find an open-source release in what we read, so we do not claim one.
Why would a lab pay when open environments exist?
Three reasons come up again and again.
Coverage. Open environments cluster around skills that are easy to grade and widely useful, such as coding and general browsing. A lab training an agent for a specific job, with specific tools, rarely finds a ready-made match, which is where the commercial builders on our companies page come in.
Contamination. Anything public may already be in a model’s pretraining data, or may have been trained on by others. A task the model has seen is a task it can answer from memory, which teaches nothing.
Rights and exclusivity. A public environment gives you no edge over anyone else using it. Labs pay heavily to avoid that: Epoch AI’s FAQ puts the markup for exclusive work at four to five times the shared price.
What should you check before training on an open environment?
Treat an open environment like any third-party code you are about to depend on.
- The license. Can you use it commercially? Can you modify and keep your changes private? Some licenses allow research use only.
- The grader. Has it been tested against wrong solutions? Look for reports of agents gaming it. If none exist, test it yourself before a long run.
- Maintenance. When was it last updated? Do issues get answered? An abandoned environment may break against new versions of the tools inside it.
- Task count and spread. Are there enough tasks, at enough levels of difficulty, to give a training signal across a full run?
- Exposure. Has the task set been public long enough that models are likely to have seen it?
Can you mix open and commercial environments?
Yes, and many teams do. A common pattern is to prototype on an open environment, find where the agent fails, then commission tasks aimed at those failures. The open environment gives a cheap baseline; the commissioned tasks give the edge. Ask any vendor whether it can deliver tasks in the format your hub or trainer already uses.
Where do open hubs sit among RL environment companies?
On our list of RL environment companies, hubs and training tools form their own group, next to the specialists and the large data providers. They are less a vendor than a starting point: free to try, quick to learn from, and a useful yardstick when a paid quote arrives.
Start with what is free, measure where it falls short, and pay only for the gap.