NeMo Gym is a library for evaluating and improving models and agents using environments. It provides infrastructure to develop environments, run evaluation and training at scale, and access a collection of popular benchmarks and training environments.
The tool addresses the need to evaluate models and agents in stateful, interactive settings where tasks require multiple steps and real-time feedback. An environment in NeMo Gym consists of a dataset defining tasks, an agent harness specifying how the model interacts with the world, a verifier that scores task completion, and state that tracks per-task execution context. This modular design allows developers to build reproducible evaluation systems that work consistently across teams and can scale to thousands of concurrent requests. The library handles the complexity of managing stateful interactions, making it suitable for scenarios like code execution, tool calling, and sandboxed environments where simple stateless scoring is insufficient.
Adopt NeMo Gym if you need reproducible evaluation across teams, require scale for multiple repeats per task or concurrent training requests, or want to seamlessly move between evaluation and agent optimization. The tool is less necessary if you only need to score model outputs with a stateless check and have no scaling requirements. NeMo Gym integrates with other environment libraries including Aviary, Harbor, OpenEnv, and Reasoning Gym, allowing you to combine benchmarks from multiple sources. It supports training with various RL frameworks and has been battle-tested in production Nemotron training.
The project maintains a substantial base of adopters who report issues from real-world use rather than the core team driving the issue tracker. Typical responses to issues or pull requests arrive within one to two weeks. Work in the issue tracker centers on documentation, core infrastructure, and community requests.
