NeMo Gym is a Python library developed by NVIDIA for evaluating and improving models and agents within interactive environments. The repository serves as infrastructure for developing environments, running evaluation and training at scale, and providing a collection of popular benchmarks and training environments. An environment in NeMo Gym consists of four core components: a dataset containing tasks to solve, an agent harness defining how the model interacts with the world, a verifier for task completion scoring, and state management for per-task execution context.
The library is designed for specific use cases where stateful evaluation is necessary. It addresses scenarios involving code execution, tool calling, and sandboxed interactions where reproducible evaluation across teams is critical. NeMo Gym enables seamless transitions between evaluation, agent optimization, and training workflows, and it scales to handle thousands of concurrent environments and requests. The library provides modular, extensible interfaces for agents, environments, tasks, and verifiers, allowing users to leverage built-in agent harnesses or integrate their own. It supports training with any reinforcement learning framework and has been battle-tested in production Nemotron training.
NeMo Gym integrates within the broader NVIDIA NeMo ecosystem for GPU-accelerated generative AI model training and agent optimization. The library seamlessly combines environments and benchmarks from other libraries including Aviary, Harbor, OpenEnv, and Reasoning Gym. It supports multiple training framework libraries such as NeMo RL, Unsloth, and VeRL. Agent harnesses are available out of the box for popular frameworks including OpenHands, Mini SWE Agent, and LangGraph, with additional harnesses like OpenCode, OpenClaw, and Pi available through recent releases.
The repository is currently in early development, with evolving APIs, incomplete documentation, and occasional bugs acknowledged by the maintainers. The most recent release, v0.4.0, introduced a unified gym CLI for discovering agents and benchmarks, BLADE diagnostics for analyzing evaluation results, agent skill evaluation capabilities, pluggable sandboxes for isolated tool-using agent execution, and hosted inference providers including Fireworks, Together.ai, and OpenRouter. The v0.3.0 release added 70+ new environments and Nemotron 3 Ultra training datasets.
According to GitGenius activity tracking, the repository shows strong community engagement with a median issue and pull request response latency of 0.0 hours and a mean latency of 57.2 hours across 592 tracked items. The most active issue labels are documentation with 175 tracked items, core-infra with 68 items, and community-request with 54 items. Primary contributors include cwing-nvidia with 647 tracked events, bxyu-nvidia with 265 events, and sephmard with 167 events. The repository shares overlapping contributors with pytorch-lightning and vllm-project/vllm, indicating cross-project collaboration within the AI infrastructure ecosystem.
NeMo Gym requires Python 3.12 or higher and runs on Linux, macOS, and Windows via WSL2 without requiring a GPU for library operation, though GPUs may be needed for specific resource servers or model inference. The library automatically installs Ray as a dependency and requires a minimum of 8 GB RAM with 16 GB recommended for larger environments. Installation uses the uv package manager, and the quickstart supports OpenAI models while also supporting local and hosted inference through vLLM, Fireworks, OpenRouter, and other providers.
