Qwen-SEA-LION-v4-32B-IT: Southeast Asian Language Model

Qwen-SEA-LION-v4-32B-IT is a 32 billion parameter instruction-tuned model developed by AI Singapore, building upon the Qwen3 architecture. It is specifically designed for the Southeast Asian (SEA) region, having undergone extensive continued pre-training on approximately 100 billion tokens from the SEA-Pile v2 corpus, which includes data from Burmese, Indonesian, Malay, Filipino, Tamil, Thai, and Vietnamese. The model was further refined through post-training on 8 million high-quality question-and-answer pairs.

Key Capabilities & Features

• Multilingual Proficiency: Enhanced understanding and generation in seven key Southeast Asian languages, in addition to English and other languages supported by Qwen3.
• Extended Context Length: Inherits a native 32,768 token context length from its Qwen3 base.
• Instruction Following: Fine-tuned for instruction-following and multi-turn chat capabilities, evaluated using SEA-IFEval and SEA-MTBench.
• Text-Exclusive: Primarily focused on text-based tasks; vision capabilities are comparable to the base Qwen3-32B model.

Use Cases & Considerations

This model is ideal for applications requiring strong language understanding and generation in a Southeast Asian context. Developers should note that the model has not been aligned for safety and requires further safety fine-tuning for production use. It may exhibit limitations such as hallucination and occasional irrelevant content, similar to many LLMs. Evaluation was conducted zero-shot across various tasks including QA, sentiment analysis, translation, and cultural knowledge using the SEA-HELM evaluation benchmark.