NeMo Gym is an open-source library from NVIDIA, published on GitHub, for evaluating and improving models and agents using interactive environments. The project has gathered around 1,300 GitHub stars and is under heavy active development in 2026.
Part of the NVIDIA NeMo platform, NeMo Gym treats an environment as a complete system — dataset, agent harness, verifier and state — and provides the infrastructure to build such environments, run evaluations reproducibly, and generate reinforcement-learning training data at scale, up to thousands of concurrent environment instances. It ships a hub of popular benchmarks and training environments across agentic, coding, knowledge, science and multilingual domains, modular interfaces for agents, tasks and verifiers, and out-of-the-box harnesses for tools such as OpenHands, Claude Code and Codex CLI. Rollouts are captured with token-level trajectories and optional OpenTelemetry tracing, rewards can be recomputed from stored rollouts without re-running inference, and the same environments plug into training frameworks including NeMo RL, VeRL and Unsloth — the stack NVIDIA uses in production Nemotron training.
NeMo Gym is free and open source under the Apache 2.0 license. You pay only for your own model inference and compute.
AI teams evaluating agents rigorously, RL researchers training models in interactive environments, and platform engineers building shared benchmark infrastructure.