Creating tree 3D models is a complex task that requires significant effort. In order to make tree modeling easier, several methods have been proposed on reconstructing tree 3D models from images or sketches. However, users need to provide appropriate images to obtain satisfactory reconstructed results, and depicting various types of trees is difficult for users who are not accustomed to drawing. Therefore, we propose a novel method to obtain tree 3D models in a zero-shot manner using text as input, by leveraging CLIP (Contrastive Language-Image Pre-training), which has garnered attention in recent years. CLIP can compute semantic similarities between input texts and images. Utilizing this property, we formulate the modeling of trees through text as an optimization problem with the evaluation by CLIP as the objective function. We adopt a genetic algorithm for the optimization problem. Since CLIP is a pre-trained model, our system does not require learning processes. The tree models are generated using the Lindenmayer System (L-system), and our method determines the parameters of the L-system that result in tree models aligned with the input text. The efficacy of our method is demonstrated through various examples.