RiNALMo is a large-scale RNA language model designed to improve the understanding of RNA structures and functions. Pre-trained on 36 million non-coding RNA sequences from various databases, RiNALMo captures hidden knowledge and structural information within RNA sequences. It achieves state-of-the-art results on several downstream tasks, particularly excelling in secondary structure prediction for unseen RNA families, where other deep learning methods struggle. The model's expressive output embeddings are used in various structural and functional tasks, demonstrating its generalization capabilities and potential for advancing RNA research. The paper details the model's architecture, pre-training process, and performance on secondary structure prediction, multi-species splice-site prediction, and mean ribosome loading prediction, highlighting its superior performance and generalization to unseen data.RiNALMo is a large-scale RNA language model designed to improve the understanding of RNA structures and functions. Pre-trained on 36 million non-coding RNA sequences from various databases, RiNALMo captures hidden knowledge and structural information within RNA sequences. It achieves state-of-the-art results on several downstream tasks, particularly excelling in secondary structure prediction for unseen RNA families, where other deep learning methods struggle. The model's expressive output embeddings are used in various structural and functional tasks, demonstrating its generalization capabilities and potential for advancing RNA research. The paper details the model's architecture, pre-training process, and performance on secondary structure prediction, multi-species splice-site prediction, and mean ribosome loading prediction, highlighting its superior performance and generalization to unseen data.