Singapore´s home-grown large language model, Sea-Lion, is seeing increasing adoption across Southeast Asia, with over 235,000 downloads to date. Major enterprises like Indonesia´s GoTo Group have already adopted the model to power their own artificial intelligence systems. Following the release of an upgraded Sea-Lion version with improved reasoning capabilities in April 2025, researchers at AI Singapore are planning to incorporate voice recognition by the end of the year, along with additional modalities such as visual recognition to further enhance usability in the linguistically diverse region.
Currently, Sea-Lion recognizes 13 languages, including Javanese, Sundanese, Malay, Tamil, Thai, Vietnamese, English, and Chinese, making it especially valuable in regions where many languages are spoken or unwritten. GoTo´s chief data officer, Mr Ofir Shalev, highlighted that using Sea-Lion as a foundation was more cost-effective than training a model from scratch. Their custom Sahabat-AI model, built on Sea-Lion, demonstrates higher accuracy for Bahasa Indonesia, Javanese, and Sundanese than comparable open-source models. The Sea-Lion project, launched in December 2023 and funded by the National Research Foundation with support from the Infocomm Media Development Authority and the Agency for Science, Technology and Research, aims to provide an open-source LLM tailored to Southeast Asia.
The latest Sea-Lion v3.5 iteration is built on Meta´s Llama 3.1 and fine-tuned for complex problem-solving and logical inference, outmatching earlier versions. Distinctive features include a 128,000-token context window for processing lengthy documents, and a ´hybrid reasoning´ mode that lets users toggle advanced capabilities to optimize resources. Enterprises such as NCS, a subsidiary of Singtel, are piloting Sea-Lion for multilingual translations, customer engagement, and regulatory monitoring, with keen interest in its scalability, integration, and security for enterprise use.
Regional applications of Sea-Lion are diverse: in Thailand, it helped a Bahasa Indonesia-speaking worker submit a labor complaint in Thai via a voice app, and it has been utilized for recognizing regional calendar systems and suggesting culturally relevant cooking ingredients. According to benchmarks developed in partnership with Stanford University´s Center for Research on Foundation Models, Sea-Lion v3.5 outperformed similar models—including ChatGPT and DeepSeek—particularly on tasks relevant to Southeast Asian languages and cultures. Dr Leslie Teo, the project lead at AI Singapore, envisions Sea-Lion serving both as a compact model for simple tasks and as a specialized companion alongside larger global models, with infrastructure support through web platforms, Telegram bots, and developer APIs to encourage broader adoption.
