Tether's AI Model Achieves Stellar Results with Genesis III Dataset
Tether's AI research division, QVAC, has made significant progress in developing artificial intelligence models. The team released Genesis III, a synthetic STEM dataset containing 191.43 billion tokens, and trained a model on it that achieved a 99.45% valid-answer rate on MMLU STEM benchmarks. This result was strong enough to earn the accompanying research paper acceptance at COLM 2026.
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, focusing on either correct answers and logic or errors and why they're wrong. The model trained on the first method alone produced a 99.45% valid-answer rate.
The full-corpus models showed even more impressive results in comparative benchmarks against Cosmopedia-v2. They improved by 28.57 points on ARC-Easy, 21.35 points on ARC-Challenge, and 15.03 points on MMLU STEM. This achievement is part of the Genesis trilogy, which started with Genesis I released on October 24, 2025, and continued with Genesis II on December 22, 2025.