A new framework, Parallel Latent Reasoning (PLR), has been introduced to improve sequential recommendation systems by exploring multiple reasoning trajectories simultaneously. This approach addresses the limitations of existing deep sequential recommendation methods by constructing parallel reasoning streams with learnable trigger tokens and employing global reasoning regularization to ensure diversity. The PLR framework adaptively combines outputs through a mixture-of-reasoning-streams aggregation. Extensive experiments have shown that the PLR framework significantly outperforms current state-of-the-art baselines while maintaining efficient inferencing. Theoretical analysis further supports its enhanced generalization capability, marking a significant advancement in the field of sequential recommendations.