Singapore’s Sea-Lion Language Model Gains Ground with Southeast Asian Focus
Sea-Lion, Singapore’s locally developed large language model (LLM), is making steady progress in Southeast Asia’s growing AI scene. With over 235,000 downloads to date, the open-source model has been adopted by regional firms such as Indonesia’s GoTo Group, highlighting its growing regional relevance.
Released by AI Singapore, Sea-Lion’s latest version—launched on April 15—introduces new reasoning capabilities. More features are already in the pipeline, with voice recognition expected in 2025, followed by visual understanding.
These upgrades aim to make the model more usable in a linguistically diverse region. Currently, it supports 13 languages, including Javanese, Sundanese, Malay, Tamil, Thai, Vietnamese, English, and Chinese.
GoTo Builds on Sea-Lion for Local AI Use
One of the earliest adopters, GoTo Group, began integrating Sea-Lion into its own AI system in February 2024. According to Chief Data Officer Ofir Shalev, the company opted for continuous pre-training rather than building from scratch—saving both time and resources.
GoTo’s Sahabat-AI model now shows improved accuracy in reading Bahasa Indonesia, Javanese, and Sundanese, outperforming peers in its class. The success illustrates how Sea-Lion’s regional grounding gives it a competitive edge over globally focused models.
Built on a $70 million investment, Sea-Lion reflects Southeast Asian linguistic and cultural contexts. According to AI Singapore, this local adaptation is what sets it apart. The model is already proving useful in tasks like document classification, insight extraction, and cross-language communication.
Also read: Singapore Airlines Partners with Salesforce to Transform Customer Service with AI
Real-World Use Cases Emerging
Sea-Lion is not just a research tool—it’s already supporting users in practical situations. In Thailand, the model helped a Bahasa-speaking migrant worker file a labour complaint in Thai. Elsewhere, it has adapted to the Thai calendar system and offered culturally relevant cooking advice, avoiding Western-biased suggestions like mayonnaise and lemon butter sauce.
Dr Ngui Jian Gang from the Sea-Lion team emphasized how the model fills cultural and linguistic gaps that larger, Western-trained models often miss.
Meanwhile, Dr Leslie Teo, AI Singapore’s senior director of AI products, said Sea-Lion v3.5 outperforms recent releases from ChatGPT and Deepseek, based on SEA-Helm—a regional benchmark developed with Stanford University.
The benchmark tests models across five dimensions: language processing, instruction-following, conversational skill, cultural accuracy, and toxicity detection in low-resource languages.
Dr Teo envisions Sea-Lion as a “companion model”—not a full replacement for GPT-4-class systems, but a useful partner in Southeast Asian contexts. “The performance is now close to the frontier,” he said, adding that the ecosystem is ready.
Sea-Lion’s features are accessible through a Telegram bot, a developer API, and an interactive Playground web tool, inviting both developers and enterprises to explore and experiment.
Growing Adoption, Growing Expectations
As Sea-Lion enters its next phase, its creators hope to attract more large-scale users. Dr Teo encouraged more adoption and feedback, noting,
Sea-Lion's regional focus, growing user base, and open-source nature position it uniquely in the race to localize AI. With continued improvements and expanding features, Singapore’s homegrown model is poised to become a vital AI engine for Southeast Asia.
