Tether’s AI research division, QVAC, has released Genesis III, a 191.43 billion-token synthetic STEM dataset spanning about 159.6 million documents across 19 domains. The corpus is divided between Option-Level reasoning, which trains models on correct answers and their underlying logic, and Failure Analysis, which teaches models from incorrect answers and their causes. A 1.7 billion-parameter model trained on the Option-Level portion achieved a 99.45% valid-answer rate on MMLU STEM benchmarks, and the accompanying paper was accepted at COLM 2026. Models trained on the full Genesis III corpus outperformed Cosmopedia-v2-trained models by 28.57 points on ARC-Easy, 21.35 points on ARC-Challenge and 15.03 points on MMLU STEM. Genesis III completes a series that began with Genesis I in October 2025 and Genesis II in December 2025. The dataset is available for non-commercial research and education under CC-BY-NC 4.0, while its model uses the commercially permissive Apache 2.0 license. QVAC says the work supports local AI tutoring and technical assistance without dependence on centralized cloud services.