aisingapore/Gemma-SEA-LION-v4-27B-IT

Gemma-SEA-LION-v4-27B-IT: Southeast Asian Language LLM

Gemma-SEA-LION-v4-27B-IT is a 27 billion parameter instruction-tuned model developed by AI Products Pillar, AI Singapore, specifically designed for the Southeast Asian (SEA) region. It is built upon the Gemma 3 architecture, inheriting its 128K context length and advanced capabilities in image and text understanding, including document comprehension, visual Q&A, and image-grounded reasoning. The model also supports advanced function calling and structured outputs for system integration.

Key Capabilities & Training

• Multilingual Proficiency: Post-trained on approximately 10 million QA samples in Burmese, English, Indonesian, Khmer, Lao, Malay, Tagalog, Tamil, Thai, and Vietnamese, making it highly proficient in these languages.
• Gemma 3 Architecture: Utilizes the Gemma 3 decoder model and its default tokenizer.
• Instruction Fine-tuning: Underwent a multi-stage post-training workflow including instruction fine-tuning, model merging, online Reinforcement Learning (RL) for instruction following and math, and on-policy alignment.
• Benchmark Performance: As of August 2025, it excels at SEA tasks compared to other open models under 200 billion parameters, with performance comparable to larger closed models, as detailed on the SEA-LION leaderboard.

Limitations

• Safety Alignment: The model has not been aligned for safety; developers must perform their own safety fine-tuning.
• Hallucination: Like many LLMs, it can hallucinate and generate irrelevant or fictional content.
• Vision Capabilities: While inheriting Gemma 3's vision capabilities, its own vision training was exclusively on the text back-end, so significant improvements in this area are not expected beyond Gemma 3 IT 27B.