Tether released QVAC Genesis III, a 191.43B-token synthetic dataset built to make small AI models far better at science, technology, engineering and math, so tutors and technical assistants can run locally instead of in the cloud.
Genesis III expands the earlier Genesis I and II corpora to 159.6M documents across 19 curriculum-aligned STEM areas, from high school to professional level. It relies on two methods: Failure Analysis, where a stronger teacher model turns a smaller student model's mistakes into new training material, and Option-Level Reasoning, which explains why the right answer is right and why every alternative is wrong. Against a token-matched Cosmopedia-v2 model it lifted scores by 28.57 points on ARC-Easy, 21.35 on ARC-Challenge and 15.03 on MMLU STEM, while a 1.7B-parameter model trained on its Option-Level data gave valid answers in 99.45% of benchmark responses.
The paper has been accepted at COLM 2026. "Most of the AI industry has focused on making models bigger," said CEO Paolo Ardoino, framing the goal as moving useful AI from cloud data centers onto the devices people already own.
Partner Offers | Get Your Instant BTC Cashback | Your Gateway To Crypto | Arcus Private Whitelist
