Kalyakan - stock.adobe.com
AI Singapore taps Alibaba Cloud to power Sea-Lion model
The fourth iteration of the regional LLM leverages Alibaba’s Qwen foundation model to improve multilingual accuracy and performance on consumer-grade hardware
AI Singapore (AISG) and Alibaba Cloud have released a new large language model (LLM) that has been improved to address the linguistic and cultural nuances of Southeast Asia.
Dubbed Qwen-Sea-Lion-v4, it combines Alibaba’s Qwen3-32B foundation model with AISG’s large regional datasets to offer an open-source alternative to Western-centric artificial intelligence (AI) models.
According to AISG, the new model has topped the Southeast Asian Holistic Evaluation of Language Models (Sea-Helm) leaderboard for open-source models with fewer than 200 billion parameters.
It is also the most recent step forward in the Sea-Lion (Southeast Asian Languages In One Network) project. Sea-Lion was first launched in 2023 to address the English bias prevalent in mainstream generative AI models.
Although models like OpenAI’s GPT-4 or Meta’s Llama series excel at English and major European languages, they frequently struggle with the low-resource languages of Southeast Asia.
In addition, global models do not account for local cultural contexts or the region’s propensity for code-switching – the practice of combining English with local vernaculars, such as Singlish in Singapore or Manglish in Malaysia.
Previous iterations of Sea-Lion have focused on creating a sovereign capability for the region, ensuring that Southeast Asian data isn’t just a footnote in the training of US-based models.
Read more about AI in ASEAN
- Malaysia’s Ryt Bank is using its own LLM and agentic AI framework to allow customers to perform banking transactions in natural language, replacing traditional menus and buttons.
- By unifying data on the Databricks platform, Vietnam’s Techcombank has built AI capabilities to deliver hyper-personalised offers to 15 million customers and expand its footprint beyond its traditional affluent base.
- Researchers at the National University of Singapore have created a wearable device that combines a camera with conversational AI powered by Meta’s Llama models to give sight to the visually impaired.
- Indonesia’s GoTo has migrated half its infrastructure to Alibaba Cloud, paving the way for AI initiatives to solve real-world business problems and support local languages.
Leslie Teo, AISG’s senior director of AI products, noted that the collaboration with Alibaba will help advance AI inclusivity and make Sea-Lion more representative of Southeast Asia. “It embodies our shared vision of accelerating AI innovation across the region and ensuring that developers, enterprises, and public institutions have access to AI that is open, affordable, and locally relevant and is designed to truly understand the languages, cultures, and communities of this region,” he said.
Qwen-Sea-Lion-v4 was built on the Qwen3-32B base model, which was pre-trained on 36 trillion tokens across 119 languages. To address the needs of the ASEAN market, AISG performed advanced post-training using over 100 billion Southeast Asian language tokens.
The model can also run on a consumer-grade laptop with 32GB of RAM. This is a critical feature for the region, where many small and medium-sized enterprises (SMEs) and developers lack access to industrial-grade GPU (graphics processing unit) clusters.
Also, instead of using a sentence-piece tokeniser, the model now uses byte-pair encoding (BPE). This makes it better at processing non-Latin scripts found in languages such as Thai and Burmese. And with a native 32k-token context length, the model can handle document-level reasoning and summarisation tasks.
Hon Keat Choong, general manager of Singapore for Alibaba Cloud Intelligence, noted that the partnership leverages the “multilingual and reasoning strengths” of the Qwen model combined with AISG's deep regional expertise.
To improve the model’s performance on colloquial speech, the teams increased the proportion of translation and cross-lingual tasks during post-training. This enables the model to better interpret informal chat and mixed-language inputs that reflect real-world usage in the region.
The model is available immediately for free download via the AISG website and Hugging Face, with four- and eight-bit quantised versions available to facilitate easier deployment.
