NEW: Tether says a 1.7B-parameter model trained on its Genesis III dataset achieved a 99.45% valid-answer rate in benchmark testing. The dataset uses failure analysis and option-level reasoning to train smaller models on STEM problems.