The University of Waterloo and other institutions have jointly open-sourced a large model tailored for mathematical problems, MAmmoTH. MAmmoTH outperforms existing open-source models on mathematical reasoning datasets, with an average accuracy improvement of 12% to 29%. Specifically, the MAmmoTH-7B model achieves an accuracy of 35% on competition-level datasets, surpassing the best open-source 7B model; the MAmmoTH-34B model reaches an accuracy of 46% on the MATH dataset, even outperforming the results of GPT-4. The performance advantage of MAmmoTH makes it the new king in the field of mathematical reasoning.