Research on natural language inference is an important task in the field of natural language processing. Traditional methods mainly rely on feature engineering, external semantic resource and tools, and the machine learning methods are combined to complete the classification of text entailment relationship. Existing deep learning methods mainly utilize deep neural network to model the sentence sequence in order to complete the representation and matching of sentence, but the following problems still exist: (1) The sentence feature representation is not rich enough; (2) The semantic expression of low-frequency words by using word vector is insufficient; (3) The problem of interactive information between sentences is ignored during modeling of sentence pair. In order to address the above three problems, from the perspective of the multi-granularity of character, word and sentence, we propose the natural language inference model with information fusion and interaction between character & word and word & sentence, and utilize deep neural network (CNN-BiLSTM) to complete the classification of text entailment relationship. Extensive experiments were conducted on the two public datasets of SNLI and MNLI. With less parameters, our model outperforms the state-of-the-art models.