Multi-Task Learning (MTL) has gained significant attention, especially in modern Recommender Systems (RS). A prevalent MTL strategy involves the Mixture-of-Experts (MoE) paradigm. However, existing MoE frameworks often use the Multi-Layer Perceptron (MLP) as an expert, lacking multi-task ability, and suffer from suboptimal performance due to simplistic gate-based feature fusion. This paper introduces a novel Task-Intensive Expert (TI-Expert) model tailored for multi-task recommendation scenarios. Our approach incorporates a well-designed hyper network for expert parameter sharing, generating task-specific parameters for each TI-Expert to produce distinct task-specific features. Additionally, we propose integrating a brainstorming layer to enhance communication among experts. Extensive experiments on several real-world datasets demonstrate that our method outperforms state-of-the-art multi-task recommendation models in terms of performance.