Universitas Terbuka (UT) provides student support services to improve student academic outcomes and student persistence for their completion in Distance Education. However, the average cumulative and semester GPA of Bachelor and Diploma programs from academic year 20161-20182 show below labor market standard GPA (2.75%). Solution to this problem is early detection on academic failure risk through the implementation of classification data mining to predict student at-risk academic failure using Online Tutorial (Tuton) activities data and student’s personal information. Classification student at-risk academic failure based on their semester GPA in order to early detect not only on the initial semester but also on the following semester. Furthermore, semester GPA has a strong positive correlation to cumulative GPA so that semester GPA is considered as an early indication of the risk of failure. The classification algorithm for student at-risk failure early detection model using naïve bayes, logistic regression, SVM, decision tree (CART, C5.0), random forest, and adaboost. The initial model testing stage use data from Tuton activities on 20182-20191. Splitting method of training data set and testing data set using Stratified K-fold in 10 times iteration and experimenting without class imbalance sampling and random undersampling method (50P:50N, 70P:30P, 66P:33P, 60P:40N) on training data set. On The initial model testing stage shows that F1-scores on fourth week are not significantly different from the eighth week so early intervention on fourth week is the right time for student to study harder on the next assignments. The highest F1-score from the initial model testing stage is without sampling imbalance on training data set using random forest (90.20%), adaboost (89.20%), and decision tree CART (88.10%). The three best algorithms will be tested again on the final testing stage using Tuton activity on 20192 as testing data set. The F1-score results on the final student at-risk of failure early detection model stage shows that adaboost algorithm highest performance (84.7%) and followed by random forest (83.8%). Based on recall results, CART showed the highest performance (99.9%) but tend to positive class overfitting so that it was no better than intervening all of students. The best performance for student at-risk of failure early detection models at UT is using adaboost algorithm.