AlienKevin/nemotron-terminal-8b-repro

TEXT GENERATIONConcurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Apr 9, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

AlienKevin/nemotron-terminal-8b-repro is an 8 billion parameter language model, a reproduction by the Marin community of NVIDIA's Nemotron-Terminal-8B. It is fine-tuned on the Nemotron-Terminal-Corpus dataset, utilizing the Qwen3-8B base model. This model specializes in terminal-based task execution, demonstrating improved performance over the original Nemotron-Terminal-8B on relevant benchmarks. It is designed for applications requiring robust command-line interaction and automation.

Loading preview...

Model Overview

AlienKevin/nemotron-terminal-8b-repro is an 8 billion parameter language model developed by the Marin community. It serves as a reproduction of NVIDIA's Nemotron-Terminal-8B, fine-tuned on a substantial portion of the Nemotron-Terminal-Corpus (366,000 examples). The model is built upon the Qwen/Qwen3-8B base model and was trained for 2 epochs with a sequence length of 32,768 tokens.

Key Capabilities & Performance

This reproduction demonstrates enhanced performance in terminal-based task execution compared to the original Nemotron-Terminal-8B. Key benchmark results include:

  • Terminal-Bench 2.0: Achieved 15.9%, a +2.9 percentage point improvement over the released Nemotron-Terminal-8B (13.0%).
  • Terminal-Bench Lite: Achieved 29.0%, a +6.0 percentage point improvement over the released Nemotron-Terminal-8B (23.0%).

These results indicate its proficiency in handling complex terminal environments and executing commands accurately.

Training Details

The model was trained using the Marin framework (Levanter-based) on TPU v5p-32 hardware. The training utilized specific subsets of the Nemotron-Terminal-Corpus, including skill_based_easy, skill_based_medium, skill_based_mixed, and diverse_complex examples.