NVIDIA researchers have tackled a persistent bottleneck in AI development: the lack of high-quality data for training agents that can navigate computer terminals. While many models excel at general conversation, using them to execute complex command-line tasks often leads to catastrophic errors or hallucinations. To bridge this gap, the team introduced Terminal-Task-Gen, a sophisticated pipeline that generates synthetic tasks based on specific skills and seed instructions. This method allows for the creation of diverse, complex scenarios that real-world datasets often lack, providing a robust foundation for specialized training.

The results of this data-centric approach are striking. The new Nemotron-Terminal family, built on the Qwen3 architecture, shows massive improvements across the board. For instance, the 32B parameter version leaped from a meager 3.4% success rate to an impressive 27.4% on the Terminal-Bench 2.0 benchmark. These models demonstrate that smart data engineering—focusing on curriculum learning and long-context training—can empower smaller models to punch far above their weight class, often matching or exceeding the capabilities of much larger, general-purpose systems.

By open-sourcing the Terminal-Corpus and model checkpoints, NVIDIA is providing the community with the tools needed to build more reliable autonomous systems. Terminal capabilities are crucial for developers and IT professionals, as they allow AI to handle server management, software installation, and complex file manipulations. This release marks a significant step toward "terminal-native" AI that understands the nuances of system shells as fluently as human language, potentially automating the most tedious parts of technical workflows.

NVIDIA researchers have tackled a persistent bottleneck in AI development: the lack of high-quality data for training agents that can navigate computer terminals. While many models excel at general conversation, using them to execute complex command-line tasks often leads to catastrophic errors or hallucinations. To bridge this gap, the team introduced Terminal-Task-Gen, a sophisticated pipeline that generates synthetic tasks based on specific skills and seed instructions. This method allows for the creation of diverse, complex scenarios that real-world datasets often lack, providing a robust foundation for specialized training.

The results of this data-centric approach are striking. The new Nemotron-Terminal family, built on the Qwen3 architecture, shows massive improvements across the board. For instance, the 32B parameter version leaped from a meager 3.4% success rate to an impressive 27.4% on the Terminal-Bench 2.0 benchmark. These models demonstrate that smart data engineering—focusing on curriculum learning and long-context training—can empower smaller models to punch far above their weight class, often matching or exceeding the capabilities of much larger, general-purpose systems.

By open-sourcing the Terminal-Corpus and model checkpoints, NVIDIA is providing the community with the tools needed to build more reliable autonomous systems. Terminal capabilities are crucial for developers and IT professionals, as they allow AI to handle server management, software installation, and complex file manipulations. This release marks a significant step toward "terminal-native" AI that understands the nuances of system shells as fluently as human language, potentially automating the most tedious parts of technical workflows.