Tether AI Research achieves 99% valid-answer rate with Genesis III model

Tether’s AI research division, operating under the name QVAC, has released Genesis III, a synthetic STEM dataset containing 191.43 billion tokens. A 1.7 billion-parameter model trained on a portion of that dataset hit a 99.45% valid-answer rate on MMLU STEM benchmarks, a result strong enough to earn the accompanying research paper acceptance at COLM 2026. What Genesis III actually is The dataset spans approximately 159.6 million documents across 19 different STEM domains, organized into three difficulty tiers: high school, college, and professional. QVAC split the corpus into two distinct training methodologies. The first, called Option-Level (OL) reasoning, trains models by focusing on correct answers and the logic behind them, accounting for 108.67 billion tokens. The second, Failure Analysis (FA), does the opposite: it trains models on errors and why they’re wrong, contributing another 82.76 billion tokens. The model trained on the OL split alone produced the headline 99.45% valid-answer rate. But the full-corpus models told an even more compelling story in comparative benchmarks. Against Cosmopedia-v2, a well-known baseline dataset, Genesis III-trained models improved by 28.57 ...