Open-source RL environments: hubs, libraries and task sets

The list

New to the term? Our RL environments explainer covers the anatomy and the training loop.

Licenses are the code licenses GitHub showed on November 9, 2026. Where GitHub could not detect one, we read the LICENSE file. A code license does not automatically cover bundled datasets, Docker images or the content of websites a replica hosts. Read those terms separately.

Hubs and frameworks

Prime Intellect Environments Hub and verifiers

Prime Intellect launched the hub on August 27, 2025 with a stated thesis: "If high-quality environments remain expensive and closed, open-source models will fall further behind" (Prime Intellect). Its companion, the MIT-licensed verifiers library, packages an environment as a task dataset, a harness and a rubric, and integrates with Prime Intellect's prime-rl trainer. Prime Intellect pays bounties to grow the catalog: $100 to $500 for open-access tasks and $1,000 to $5,000 or more for application-only tasks, paid on completion. Benchmark ports count as complete only when they reproduce known scores (October 2025). Bounty tasks reported by Epoch include benchmark ports, MCP-server interaction, nanoGPT speedrunning and textbook questions.

OpenEnv

OpenEnv, from Meta-PyTorch and Hugging Face, is a spec more than a catalog. Environment authors implement step(), reset() and close(), and the environment runs as a "secure, semantically clear" sandbox usable for training and deployment (Hugging Face, October 2025). Version 0.1 was released as an RFC. The draw is trainer integrations: TorchForge, TRL, verl, SkyRL and Unsloth.

NVIDIA NeMo Gym

NeMo Gym runs environments as servers separate from the training loop and records trajectories in the OpenAI Responses format (NVIDIA docs). The core is Apache-2.0, while some bundled data carries Creative Commons terms. NVIDIA labels it early development, so pin versions.

Gymnasium

For classic control and game RL, Gymnasium is still the standard. It is not built for text actions or container sandboxes. Our RL gym guide explains the API and how the LLM-era frameworks borrow from it.

Task sets by skill

• Coding: SWE-Gym for training, SWE-smith and R2E-Gym for generating more tasks, SWE-bench for evaluation. DeepSeek reportedly used 24,667 coding tasks extracted from GitHub (SemiAnalysis), which shows the scale a frontier coding run uses.
• Terminal: Terminal-Bench 2.0. At publication in January 2026, frontier agents solved under 65% of its 89 tasks.
• Web: WebArena for self-hosted sites; BrowserGym to run WebArena, WorkArena and others behind one API.
• Desktop: OSWorld. Humans scored 72.36% and the best model 12.24% at release in April 2024.
• Enterprise: WorkArena on ServiceNow and TheAgentCompany, where the best agent finished 30% of tasks at release.
• Tool use with a user: tau-bench and tau2-bench. tau-bench introduced pass^k; gpt-4o scored under 50% and its retail pass^8 was under 25%.
• Math and verifiable reasoning: the Tulu 3 RLVR data.

When open source is enough

• Algorithm research. If the question is about the optimizer, the reward shaping or the infrastructure, a public environment makes your results reproducible and comparable.
• Open-weight post-training on common skills. Coding and math have the deepest public pools.
• Building your stack. Wire up the trainer, rollout workers and sandboxing on free environments before you spend on tasks.
• Calibrating vendors. Run your checkpoint on public sets first. You then know what pass rate a vendor's "hard" tasks must beat. Our evaluation checklist covers the rest.

When labs buy instead

• The domain is not public. Enterprise replicas of Slack, Salesforce or SAP, and workflows in finance or law, have few open equivalents. Vendors argue domain coverage is "the binding constraint" for in-house teams (Invisible, a vendor with an interest in that view).
• Public tasks are contaminated or saturated. OpenAI stopped reporting SWE-bench Verified in February 2026, citing flawed tests and training exposure (OpenAI). a16z cited OSWorld scores of 72.7% for Claude Opus 4.6 and 75% for GPT-5.4 in March 2026, about the human baseline (a16z).
• Secrecy. Epoch's sources cite keeping training priorities confidential as a reason labs build in-house. The same logic drives exclusive purchases. Most vendors close source their environments and sell them under exclusive contracts; SemiAnalysis names Prime Intellect as the exception.
• Quality and volume. Public sets were built as benchmarks. Training needs many tasks inside a pass-rate band on your model, with graders hardened against your model's hacks.

Big labs both build and buy, per Epoch AI and TechCrunch. A workable split: open frameworks for the plumbing, bought or in-house tasks for the frontier skills. Prices for the bought part are on our pricing page, terms are in the glossary, and the commercial vendors are compared on our home page.