The stiffness of a hand joint is correlated to the co-contraction of the agonist and antagonist muscles surrounding the joint. First step towards using co-contraction and other hand motions for a control command of a robot or prosthesis is to reliably recognize and distinguish them. In this paper we implement, train and validate five Recurrent Neural Network (RNN) models using Long-Short Term Memory (LSTM) layers to develop a high-quality classifier of three hand gestures: flexion, extension and co-contraction intended to command closing, opening and stiffness modulation of the robot, respectively. 16-channel surface electromyography (sEMG) signals were recorded using a MaxSens system to form three datasets, two of which contain the downsampled data of the third, original dataset. The performance of the models was similar and sufficiently high while using all three datasets: F1-score values were over 98%. However, the model containing three LSTM and three fully connected layers stood out from the others for its inference time and training stability. The best-performing candidate was further tested with a reduced number of EMG channels and reached F1-score equal to 94.31% for the original dataset, 89.71% and 93.68% for two downsampled datasets when 2 EMG channels were used.