aisingapore/Qwen-SEA-LION-v4-32B-IT

Qwen-SEA-LION-v4-32B-IT Overview

Qwen-SEA-LION-v4-32B-IT is an instruction-tuned large language model developed by AI Singapore, building upon the Qwen3-32B foundation. This model 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. This corpus includes data across seven key SEA languages: Burmese, Indonesian, Malay, Filipino, Tamil, Thai, and Vietnamese, alongside English.

Key Capabilities & Features

• Multilingual Proficiency: Enhanced understanding and generation in Burmese, English, Indonesian, Khmer, Lao, Malay, Mandarin, Tagalog, Tamil, Thai, and Vietnamese.
• Qwen3 Architecture: Inherits robust reasoning capabilities and support for over 100 languages from its Qwen3 base.
• Extended Context Length: Features a native context length of 32,768 tokens.
• Instruction Following: Post-trained on 8 million high-quality question-and-answer pairs for improved instruction adherence.
• Evaluation: Assessed using SEA-HELM for general language tasks (QA, Sentiment, Translation, MMLU Lite) and SEA-IFEval/SEA-MTBench for instruction-following and multi-turn chat, with results available on the SEA-LION leaderboard.

Considerations for Use

Developers should note that the model has not been aligned for safety and requires custom safety fine-tuning. It may exhibit typical LLM limitations such as hallucination and occasional irrelevant content generation. The model supports a "thinking mode" which can be enabled via the enable_thinking=True parameter during generation.