AI Singapore launches Qwen SEA LION v4
Quantised 4-bit and 8-bit versions are also available.
AI Singapore has launched Qwen-SEA-LION-v4, a new regional language model built on Alibaba’s Qwen3-32B framework.
According to the joint announcement, the model incorporates additional Southeast Asia-specific training to address local linguistic, cultural, and commercial use cases.
AI Singapore said the model currently ranks first on the SEA-HELM leaderboard among open-source systems with fewer than 200 billion parameters.
The groups reported that they were further trained on more than 100 billion tokens from Southeast Asian languages to enhance their ability to handle local expressions and mixed English–regional usage.
The organisations stated that the model is efficient enough to run on consumer hardware equipped with 32 GB of RAM.
Quantised 4-bit and 8-bit versions are also available. Qwen-SEA-LION-v4 supports a native 32,000-token context window and now uses a BPE tokenizer to improve multilingual processing.
The broader Qwen3 family was pre-trained on 36 trillion tokens spanning 119 languages and dialects. Post-training steps were added to strengthen translation and cross-lingual performance, especially for real-world code-switched inputs.
Alibaba Cloud provided the base model and post-training support, while AI Singapore handled regional data curation, optimisation, and evaluation for open-source release. The model is available for free download through AI Singapore and Hugging Face.