Multi-UAV (Unmanned Aerial Vehicle) cooperative task planning has demonstrated great potential in dealing with complex and dynamic environments. However, most existing multi-UAV task planning methods rely on static modeling under the assumption of homogeneous UAVs, along with manually designed heuristic algorithms or solvers. These methods struggle to adapt to real-time changes in task demands and environmental conditions. Heterogeneous UAV—with differences in capabilities, endurance—introduce further challenges in coordination and task allocation. To address these challenges, this paper proposes a dynamic heterogeneous multi-UAV task planning method based on reinforcement learning, aiming to overcome the limitations of traditional methods. Specifically, a sequential Markov decision process is formulated to model the dynamic heterogeneous multi-UAV task planning problem, and a dual-encoder lightweight single-decoder policy network is designed, incorporating self-attention and cross-attention mechanisms. Experiment results demonstrate that the proposed method outperforms widely-used solvers in both solution quality and computational speed, verifying its effectiveness and superiority.