RiNALMo is a large RNA language model with 650 million parameters, pre-trained on 36 million non-coding RNA sequences from multiple databases. It is designed to extract hidden knowledge and capture structural information from RNA sequences. RiNALMo achieves state-of-the-art results on various downstream tasks, including secondary structure prediction, multi-species splice-site prediction, and mean ribosome loading prediction. It demonstrates strong generalization capabilities, outperforming other deep learning methods in tasks where RNA families are not seen during training. The model's output embeddings provide powerful sequence representations that improve performance on structural and functional RNA tasks. RiNALMo is publicly available on GitHub and has been tested on multiple RNA-related tasks, showing its effectiveness in RNA structure and function prediction. The model's architecture includes a BERT-style Transformer encoder with advanced techniques such as RoPE, SwiGLU activation, and FlashAttention-2. It was pre-trained using masked language modeling (MLM) and fine-tuned for various downstream tasks. The results show that RiNALMo generalizes well across different RNA families and outperforms other RNA language models and deep learning methods on most datasets. The model's performance is evaluated on multiple tasks, including secondary structure prediction, splice-site prediction, and ribosome loading prediction, demonstrating its versatility and effectiveness in RNA-related applications.RiNALMo is a large RNA language model with 650 million parameters, pre-trained on 36 million non-coding RNA sequences from multiple databases. It is designed to extract hidden knowledge and capture structural information from RNA sequences. RiNALMo achieves state-of-the-art results on various downstream tasks, including secondary structure prediction, multi-species splice-site prediction, and mean ribosome loading prediction. It demonstrates strong generalization capabilities, outperforming other deep learning methods in tasks where RNA families are not seen during training. The model's output embeddings provide powerful sequence representations that improve performance on structural and functional RNA tasks. RiNALMo is publicly available on GitHub and has been tested on multiple RNA-related tasks, showing its effectiveness in RNA structure and function prediction. The model's architecture includes a BERT-style Transformer encoder with advanced techniques such as RoPE, SwiGLU activation, and FlashAttention-2. It was pre-trained using masked language modeling (MLM) and fine-tuned for various downstream tasks. The results show that RiNALMo generalizes well across different RNA families and outperforms other RNA language models and deep learning methods on most datasets. The model's performance is evaluated on multiple tasks, including secondary structure prediction, splice-site prediction, and ribosome loading prediction, demonstrating its versatility and effectiveness in RNA-related applications.