Fast-ThinkAct is a new framework designed for Vision-Language-Action (VLA) tasks. It focuses on improving reasoning efficiency by compressing lengthy textual reasoning into more concise latent chain-of-thoughts (CoTs). This method allows for significant speed improvements, achieving up to a 9.3 times faster inference while still delivering strong reasoning performance for tasks requiring interaction with dynamic visual environments. The framework operates by training a teacher model that uses detailed textual reasoning combined with action-aligned visual rewards. A student model is then distilled from this teacher, learning to perform compact latent reasoning. This process is guided by a preference-driven objective that helps align the learned manipulation trajectories with both language and visual planning. As a result, Fast-ThinkAct enhances policy learning, connecting concise reasoning directly to action execution. In extensive experiments, Fast-ThinkAct demonstrated a remarkable reduction in inference latency, decreasing it by up to 89.3% compared to existing state-of-the-art VLA models. It maintained effective long-horizon planning and adaptability even in complex manipulation scenarios, as shown through evaluations on multiple benchmarks like SimplerEnv and RoboTwin2.0. The qualitative results indicate that the verbalized reasoning from the student model is not only more compact but also more accurate than the verbose outputs from the teacher model, highlighting its practical advantages in real-world applications.