With the rapid development of artificial intelligence (AI) technology, large language models (LLMs) have made significant progress in natural language generation and understanding. However, LLMs still face challenges when dealing with non-English languages, especially those with limited resources. Although the emergence of generative multilingual LLMs has provided a solution to this problem, the language coverage still needs to be expanded. Recently, the introduction of the XLM-R self-encoding model has become a significant milestone. This model has 278 million parameters and covers a range of languages from 100 to 534, including 534 languages from the Glot500-c corpus, which is a blessing for low-resource languages. In addition, effective strategies to address data scarcity include vocabulary expansion and continuous pre-training. The tremendous success of these models has sparked more research interest. One research team is dedicated to addressing the limitations of small model sizes and aims to expand the capabilities of LLMs to cover a wider range of languages. They discuss language adaptation strategies for expanding model parameters to 10 billion LLMs to improve contextual and linguistic relevance between different languages. However, adapting LLMs to low-resource languages still faces some challenges, such as data sparsity, domain-specific vocabulary, and language variations. To address these issues, the team proposes solutions such as vocabulary expansion, continued training of open LLMs, and the use of LoRA low-rank heavy parameterization as adaptation strategies. A team of researchers from the University of Munich, Munich Center for Machine Learning, University of Helsinki, Lisbon Institute of Technology (Lisbon ELLIS Unit), Telecom Institute, and Unbabel has proposed a new large language model called MaLA-500. This model is designed to cover a wide range of 534 languages and incorporates vocabulary expansion and continuous LLaMA 2 pre-training using Glot500-c. By analyzing the SIB-200 dataset, the results show that MaLA-500 performs better than currently available open LLMs on a relatively or slightly larger model size. The model has achieved remarkable results in contextual learning, demonstrating its adaptability and importance in different language environments. MaLA-500 provides a good solution to the current problem of LLMs not supporting low-resource languages. Through methods such as vocabulary expansion and continuous pre-training, it demonstrates state-of-the-art results in contextual learning. Vocabulary expansion aims to expand the model's vocabulary to cover a wider range of languages, enabling it to understand and generate content in various languages. In conclusion, this research is crucial for improving the accessibility of language learning modules (LLMs) and making them useful for various specific language use cases, especially for low-resource languages. With the continuous advancement of technology, we can expect to see more innovative solutions to overcome the challenges faced by LLMs in handling non-English languages. This will further drive the application and development of artificial intelligence in cross-language communication and understanding.