May 6, 2025
SINGAPORE – Singapore’s home-grown large language model (LLM), Sea-Lion, is steadily gaining traction, with some 235,000 downloads so far, bolstered by adoption by large companies such as GoTo Group in Indonesia.
After releasing the latest model with “reasoning” capabilities on April 15, its researchers at AI Singapore told The Straits Times that they plan to add voice recognition later in 2025, followed by other modalities such as visual recognition.
The new features are expected to enhance the model’s usability in a region rich in spoken and unwritten languages. The model currently recognises 13 languages, including Javanese, Sudanese, Malay, Tamil, Thai and Vietnamese, as well as English and Chinese.
Sea-Lion is already tapped by some businesses for its language features, with GoTo among the first enterprises to adopt Sea-Lion in February 2024 as a base to build its own artificial intelligence (AI) system.
Its chief data officer, Mr Ofir Shalev, noted: “Training a model from scratch is often prohibitively expensive. So like many in the industry, we adopted a continuous pre-training approach – building on an existing model as the starting point.”
GoTo’s Sahabat-AI model is now benchmarked as more accurate in reading and interpreting Bahasa Indonesia, Javanese and Sundanese than other models of similar size, according to Mr Shalev.
The $70 million Sea-Lion initiative to build an open-source LLM that reflects the native characteristics of South-east Asia was publicly launched in December 2023.
It is funded by the National Research Foundation and backed by the Infocomm Media Development Authority and Agency for Science, Technology and Research.
Sea-Lion’s latest iteration v3.5, built on another open-source model, Llama 3.1, is fine-tuned to be more adept at complex problem-solving, logical inference and multi-step instructions than earlier versions.
It is one of the few models worldwide to offer a “hybrid reasoning” mode, allowing users to toggle advanced reasoning on or off – saving time and computing resources for straightforward tasks.
A key improvement is the model’s 128,000-token context window, enabling it to process and understand much longer documents and conversations without losing track of earlier information. This is on a par with leading models such as GPT-4o and Meta’s Llama 3.1, and surpassed by only a few, such as Google’s Gemini and Anthropic’s Claude.
The chief scientist at Singtel technology subsidiary NCS, Mr Ying Shaowei, calls Sea-Lion v3.5 a significant advancement for Singapore’s AI ecosystem.
NCS is expanding its Sea-Lion pilot to support legal and compliance document translation. It will use the latest model to also perform multilingual customer engagement and cross-border regulatory change detection, and to unify internal content across languages and formats.
Mr Ying said: “We are deeply interested in how well the model performs as part of a larger system, how it integrates with existing enterprise workflows, its interoperability, its total cost of deployment and its security posture.”
These critical factors will determine if Sea-Lion can be scaled up in real operational environments, Mr Ying said.
He added: “Sea-Lion is showing real promise on several of these fronts, and we are continuing to test its role in more complex AI-driven solutions that go beyond language to include insight extraction, document classification and voice-based interaction.”
In Thailand, Sea-Lion has been adopted into a voice app that, in one instance, helped a Bahasa Indonesia-speaking worker file a complaint with the Labour Department to recoup her unpaid salary. The complaint was filed in Thai.
Its other use cases include a Python programming script that recognises the unique Thai calendar system of adding 543 years to the Gregorian year. It also recommended Asian condiments for cooking to a Filipino-speaking user.
Dr Ngui Jian Gang, who demonstrated the model to The Straits Times, said: “GPT 4 recommends mayonnaise, which is not very local, and also lemon butter sauce, which is delicious, but also not very local.”
Dr Leslie Teo, senior director of AI products and the Sea-Lion team’s lead at AI Singapore, said when evaluated against an industry benchmark tailored for the region, Sea-Lion v3.5 outperformed ChatGPT’s and Deepseek’s recent models.
The benchmark called SEA-Helm – the South-east Asian Holistic Evaluation of Language Models – is developed by AI Singapore in partnership with Stanford University’s Centre for Research on Foundation Models.
The evaluation was done across five metrics – comprising natural language processing, instruction-following, conversational ability, linguistic and cultural performance, and toxicity detection for low-resource South-east Asian languages.
Dr Teo describes Sea-Lion as best used as a “small model” for simple tasks, or as a “companion model” paired with large models such as ChatGPT, Claude or DeepSeek to fill gaps in the South-east Asian context. He hopes that it will be useful for organisations with operations in the region.
With the new and planned improvements, he is hopeful of drawing more adopters. He said: “What has changed this time is that performance is close to the frontier.”
The Sea-Lion ecosystem is also ready, he added.
An interactive web platform called Playground lets users try out the model; its Telegram bot allows users to engage it in their preferred language; and it has application programming interfaces that enable developers and organisations to integrate Sea-Lion into their applications and workflows.
Dr Teo said: “We feel that the model is good enough. We want people to use it more. We want to get to a point where we have real big users using it, criticising it and helping us make it better.”