AISG Abandoning Meta Models and Shifting to Alibaba's Tongyi Qianwen Architecture

The Singapore National AI Programme (AISG) recently announced a major strategic adjustment, abandoning Meta's Llama series models in its latest Southeast Asian language large model project and fully shifting to Alibaba's Tongyi Qianwen Qwen open-source architecture. This decision not only marks a critical expansion of Chinese open-source AI models in global influence but also directly addresses technical pain points in Southeast Asia's multilingual scenarios.

According to official announcements from AISG, its "Qwen-SEA-LION-v4" model, officially launched on November 25, quickly topped authoritative open-source leaderboards measuring Southeast Asian language capabilities (such as XNLI and TyDi QA), significantly outperforming previous models based on the Llama series in tasks like semantic understanding and text generation for regional languages including Indonesian, Thai, and Malay.

The core driver of this adjustment is to address the long-standing "language adaptation challenge" that has plagued localized AI development. Previously, international mainstream open-source models represented by Meta's Llama, while having a certain foundation in multilingual support, exhibited significant shortcomings when handling Southeast Asian minority languages—issues such as incomplete vocabulary coverage, grammatical comprehension deviations, and insufficient cultural context adaptation led to low efficiency in deploying AI applications in vertical fields like legal consultation and government services. For example, in the scenario of generating Indonesian legal documents, the Llama model once had a term misuse rate as high as 30%, severely impacting user experience.

An AISG technical lead stated: "Southeast Asia has a population of over 600 million, encompassing more than 10 official languages and hundreds of dialects. Linguistic diversity is a natural barrier to AI deployment. Qwen-SEA-LION-v4 achieves a leap from 'general translation' to 'contextualized understanding' by deeply optimizing regional language datasets and integrating Alibaba's multilingual experience accumulated in scenarios like e-commerce and cross-border services."

Notably, this collaboration is not a simple technology procurement but a deep synergy based on the open-source ecosystem. The AISG team and Alibaba DAMO Academy jointly optimized the attention mechanism of the Qwen model, conducted specialized training targeting the agglutinative features of Southeast Asian languages (such as complex affixes in Malay), and opened some fine-tuning interfaces for local developers to customize. This "global technology + regional demand" model is regarded as a typical case of Chinese open-source AI "going global."

Industry analysts pointed out that as a hub for AI R&D in Southeast Asia, Singapore's technology selection carries symbolic significance. This shift not only validates the technological leadership of Tongyi Qianwen in complex language scenarios but also reflects the continuous rise in global developers' recognition of Chinese open-source models. Currently, the Qwen series models have garnered over 500,000 stars on GitHub, with downloads exceeding 10 million times, covering multiple sectors such as finance, healthcare, and education.

With the deployment of "Qwen-SEA-LION-v4," Singapore plans to launch over 10 Southeast Asian language AI applications by 2026, covering scenarios such as cross-border trade, tourism services, and public education. AISG stated that it will continue to deepen cooperation with Chinese tech enterprises like Alibaba in the future to explore more innovative possibilities of "technology + scenario."