NVIDIA has introduced Nemotron-Terminal, a family of models designed for autonomous Linux terminal operations, including dependency installation, coding, debugging, and end-to-end engineering tasks without human intervention. Built on Qwen3 and trained on the specially curated Terminal-Corpus dataset, these models demonstrate significant performance improvements on Terminal-Bench 2.0, with the largest 32B parameter model outperforming much larger models like Qwen3-Coder 480B and GPT-5-Mini.

The dataset and models are openly available under NVIDIA Open Model License and CC-BY-4.0 License, respectively, with all resources released on Hugging Face.

Read more in the arXiv paper and access the models on Hugging Face.