The stablecoin giant's AI division drops 191 billion tokens of STEM training data designed to make small language models actually think
Tether’s AI research arm, QVAC, has released Genesis III, a massive dataset containing 191.43 billion tokens aimed at teaching smaller AI models how to reason through STEM problems rather than just guess at patterns. The data covers 19 distinct STEM domains and is freely available on Hugging Face under a Creative Commons CC-BY-NC 4.0 license.
What Genesis III actually does
The core innovation here is what QVAC calls a “dual teacher-distillation strategy.” In plain terms, the system uses a weaker AI model as a student, then learns from that student’s mistakes and successes in two different ways.
When the student model gets an answer wrong, the system generates corrective explanations that walk through why the error happened and what the right reasoning looks like. When the student gets something right, the system creates contrastive reasoning that compares the correct answer against other plausible options, explaining why each alternative falls short.
The numbers tell a compelling story
Genesis III is a significant escalation from its predecessors. Genesis I contained 41 billion tokens, Genesis II scaled up to 148 billion, and now Genesis III lands at 191.43 billion.
Performance benchmarks suggest the scaling is paying off. Models trained exclusively on Genesis III data achieved up to 28.57% performance gains on the ARC-Easy benchmark and 21.35% improvements on ARC-Challenge compared to models using existing open datasets like Cosmopedia-v2. The models also demonstrated valid answer rates as high as 99.45%. Additional evaluations on GPQA Diamond and MMLU STEM subsets further validated the dataset’s effectiveness.
A research paper accompanying the release was submitted to arXiv on September 17, 2026, providing the technical documentation that lets independent researchers verify these claims.
Why a stablecoin company is building AI datasets
The Genesis series reflects a broader strategic bet on decentralized, on-device AI, systems that can run on edge devices without needing to phone home to massive cloud servers. The dataset spans various difficulty levels and educational environments across its 19 STEM domains.
By releasing the data under a Creative Commons license and hosting it on Hugging Face, QVAC is making a deliberate play for community adoption. The dataset is available under the identifier qvac/GenesisIII, and anyone can download it without cost.
