H2O-Danube-1.8B is a 1.8B parameter language model trained on 1T tokens, adhering to the core principles of LLaMA 2 and Mistral. Despite using a significantly smaller total token count during training compared to similar-sized reference models, it demonstrates highly competitive performance on multiple benchmark tests. Furthermore, we have released a fine-tuned chat model that has undergone both supervised fine-tuning and direct preference optimization training. H2O-Danube-1.8B is open-sourced under the Apache 2.0 license, further democratizing access to large language models and enabling a wider audience to benefit economically.