We show that base models possess fundamental reasoning capabilities that can be activated through appropriate steering vectors, recovering up to 91% of the performance gap to thinking models.