Temporal Sentence Grounding in Videos (TSGV) represents an key technology for interpreting unstructured video content through the application of natural language descriptions. Amid the rapid evolution of large-scale models, their representation capability has notably improved downstream tasks, particularly within TSGV. However, existing work often relies on task-specific model architectures and training objectives when utilize large models, which restricts their scalability and transferability. In this work, we propose a novel Mask Time as Language (MTL) model, which is based on the language model without any task-specific “heads”, and we develop a matched training strategy with time tokenization and mask language modeling methods, aligning it with large models. Moreover, to resolve lack of video-text datasets annotated with time, we present a novel dataset expansion approach in the TSGV task that combines short video clips with text descriptions in existing datasets to synthesize longer videos with temporal annotation, using larger data scale to improve model performance. Experiments demonstrate that the proposed MTL model excels on Charades-STA and ActivityNet-Captions datasets, while also showing the potential of transferring large pre-trained models into the TSGV task.