Gemma-SEA-LION-v4-27B-IT
Gemma-SEA-LION-v4-27B-IT is the instruction-tuned release of AI Singapore's Gemma 3-based 27B checkpoint. Its regional post-training uses about 10 million question-and-answer samples. It retains image input capabilities, although this adaptation trains the text components.
How this release works
The card expects vision performance to remain comparable to Gemma 3 27B IT because regional training focuses on text. It lists Burmese, English, Indonesian, Khmer, Lao, Malay, Mandarin, Tagalog, Tamil, Thai and Vietnamese. The later VL release adds separate image-text training.Gemma-SEA-LION-v4-27B-IT official model card
How to assess it for your work
Practical guidance. Use the IT and VL pair to isolate the value of regional visual training. Give both the same document image and the same question. Check whether an error comes from reading the image, interpreting a local term, or composing the answer. This is more useful than comparing only overall response style.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | aisingapore/Gemma-SEA-LION-v4-27B-IT |
| Parent checkpoint | Gemma-SEA-LION-v4-27BGemma-SEA-LION-v4-27B-IT official model card |
| Context stated by the card | 128K tokensGemma-SEA-LION-v4-27B-IT official model card |
| License declared by the model card | MITGemma-SEA-LION-v4-27B-IT official model card |
| Model family size | 27BGemma-SEA-LION-v4-27B-IT official model card |
Related models
- Gemma-SEA-LION-v4-27B-VL
AI Singapore · SEA-LION v4
A 27B Gemma-based vision-language model trained on regional image-text instructions with SEACrowd.
Launch date not confirmedDownloadable - Qwen-SEA-LION-v4.5-27B-IT
AI Singapore · SEA-LION v4.5
A 27B multimodal model based on Qwen3.6, distilled for Southeast Asian language tasks.
Launch date not confirmedDownloadable