SG program selects Alibaba’s Qwen for SEA language AI model

AI Singapore has selected Alibaba’s Qwen3-32B foundation model to develop its latest large language model focused on Southeast Asian languages.

AI Singapore is a government-backed program that aims to boost AI adoption in Singapore.

The new model, named Qwen-SEA-LION-v4, is designed to address regional linguistic and cultural needs.

An earlier version of the SEA-LION model was based on Meta’s Llama, while another multimodal model released in August used Google DeepMind’s Gemma.

The adoption of Qwen highlights growing competition between Chinese and US open-source AI models in neutral markets like Singapore.

Alibaba said its Qwen series, launched in August 2023, has been widely adopted by developers globally.

The latest model was trained on over 100 billion tokens from Southeast Asian languages and currently tops the SEA-HELM leaderboard for open-source models under 200 billion parameters.

🔗 Source: South China Morning Post

🧠 Food for thought

Implications, context, and why it matters.

Open weights confirmed; licensing details vary by base model

Qwen-SEA-LION-v4-32B-IT is on Hugging Face (an AI model hub) 1 and Ollama (a local model runner for desktops and laptops) 2. Free 4- and 8-bit quantized downloads 3 confirm open weights (downloadable model parameters for local use plus fine-tuning).
It inherits Qwen3-32B’s architecture (the network design) and tokenizer (the component that splits text into tokens) 1. The model card (the model’s official documentation) lacks a license 4. Terms can vary by base model.
Laptops with 32GB RAM can run it 3. A 4-bit build keeps loss under 0.3% 5, which helps SMEs without industrial GPUs use it locally.
The tokenizer uses byte-pair encoding (BPE, a subword tokenization method) over sentence-piece (an alternative tokenization algorithm) 3. This improves Thai and Burmese handling for Southeast Asian language work.

Cloud vendors can target enterprises adopting SEA-LION for regional operations

Partners span AWS and Google, plus IBM, NCS, NVIDIA 6. The model runs on Google Cloud, AWS, Kaggle (a data science platform) and NVIDIA Inference Microservices (NIM) 6. This opens paths for managed options.
GoTo Group in Indonesia used an earlier SEA-LION in February 2024 7. Continuous pre-training powered Sahabat-AI with better accuracy in Bahasa Indonesia, Javanese, and Sundanese 7. Teams apply it to legal document translation, multilingual customer engagement, and regulatory change detection 7. Over 700 developers build with SEA-LION v4 in the Pan-SEA Challenge (a regional developer challenge) 6.