SEA-LION v4: Our First Multimodal Release, Open, Powerful & Small
We’re proud to introduce SEA-LION v4 – our open, most powerful and efficient, multimodal & multilingual model yet, designed to run even on a laptop. This release expands SEA-LION beyond text to support image + text inputs, while staying true to our focus on Southeast Asian languages, culture, and use cases. Built on Gemma 3 (27B), Google’s open model family for multimodal workloads, SEA-LION v4 comes with long-context capabilities, broad multilingual coverage, and modern developer features such as function calling.
What’s new in v4
- Multimodal understanding (image + text). SEA-LION v4 can take images alongside text and generate text outputs, enabling document understanding, visual Q&A, and image-grounded assistance in regional contexts.
- Long-context prompts (up to 128k tokens). Useful for multi-page materials and richer conversations. Actual usable length depends on task and deployment settings.
- Function calling support. Structured outputs let developers connect SEA-LION v4 to tools and workflows for agentic applications.
- Broader language coverage. Gemma 3 supports 140+ languages; we extend and adapt this to strengthen SEA-LION’s capabilities for Southeast Asian languages through SEA-focused post-training.
SEA-LION v4
| Gemma-SEA-LION-v4-27B-IT | |
|---|---|
| Architecture | Based on Gemma 3 27B IT |
| Parameters | 27 billion |
| Context Length | 128K |
| Performance | Outperforms models of similar size in SEA tasks, with developer-friendly features for easy adoptions |
SEA-LION v4 (Quantized models)
Quantized variants of Gemma-SEA-LION-v4-IT are available:
- Gemma-SEA-LION-v4-IT-GGUF
- Gemma-SEA-LION-NVFP4
- Gemma-SEA-LION-FP8-Dynamic
In particular, the FP8-Dynamic and NVFP4 variants have little degradation (<0.5% on average) in performance compared to Gemma-SEA-LION-v4-27B-IT, and they can be run on a laptop with vLLM and CUDA-compatible GPU.
Qualitative updates for SEA-LION v4
Technical Specifications
Training details
The dataset comprises Burmese, English, Mandarin, Indonesian, Khmer, Lao, Malay, Tagalog, Tamil, Thai and Vietnamese languages, collected from a mixture of sources including web data, code, open-source datasets, and synthetically generated datasets, amounting to a total of 500 billion tokens.
Training Procedure
We perform post-training using a variety of Reinforcement Learning (RL) methods. The instruction fine-tuning dataset combines our SEA-Instruct, Infinity-Instruct, and OpenMath-Instruct 2 with open-source datasets such as nvidia/Llama-Nemotron-Post-Training-Dataset (RL set) and zwhe99/DeepMath-103K. Prompt sampling is guided by a gradient-based analysis process.
Our post-training workflow consists of multiple stages: instruction fine-tuning, model merging, online RL for both instruction following and math using DRGPPO, and on-policy alignment via APO. For alignment, rejected-chosen pairs are generated from the target model, with the “chosen” responses obtained by rewriting and improving upon the rejected outputs.
Multilingual Proficiency
Gemma 3 supports 140+ languages out of the box, with significantly improved multilingual embeddings and cross-lingual transfer compared to earlier Gemma families.
SEA-LION v4 builds on this by specialising the multilingual backbone for SEA languages (Indonesian, Thai, Vietnamese, Tagalog, Burmese, Khmer, etc.), improving fluency.
This means Gemma 3 provides broad coverage, and SEA-LION v4 adds depth and nuance for regional languages that are underrepresented.
Evaluation Metrics
SEA-LION v4 have been rigorously evaluated using both English and Southeast Asian benchmarks:
- English Evaluation: Utilises tasks like BBH, GPQA, IFEval, Math-Hard, MMLU-Pro, and MUSR.
- Southeast Asian Evaluation: Employs SEA-HELM, which covers languages such as Burmese, Filipino, Indonesian, Malay, Tamil, Thai, Vietnamese, and includes tasks such as summarisation, toxicity detection, SEA-IFEval, SEA-MTBench, and more.
The evaluation was done zero-shot with native prompts on a sample of 100-1000 instances for each dataset.
Based on SEA-HELM’s holistic evaluation suite tailored for the SEA region:
SEA-LION v4 ranks #5 out of 55 models on our SEA-HELM benchmark – and #1 among open models under 200B parameters – outperforming much larger systems while running on a laptop with 32GB RAM.
See our leaderboard here for details.
Accessibility and Availability
SEA-LION v4 is open-source and freely available for research and commercial use. Developers and enterprises can immediately access the models on platforms such as Hugging Face, Google Cloud Vertex, AWS Sagemaker, and Nvidia NIMs (coming soon) with availability on Kaggle and Ollama rolling out in the coming days.
Hugging Face
Gemma-SEA-LION-v4-27B-IT-NVFP4
Gemma-SEA-LION-v4-27B-IT-FP8-Dynamic
Other updates:
Along with SEA-LION v4, we are also thrilled to share an update:
- Launch of Pan-SEA AI Developer Challenge
- Launched on 4 Aug 2025, we have received 700+ participants from all over the world to develop solutions for SEA. Stay tuned to the announcement of our winners, and we are excited to see what solutions can be developed with SEA-LION v4!
Conclusion:
The SEA-LION team is continuously looking to improve and uplift the AI community for SEA. We encourage users to try out our models and are more than happy to receive feedback, which will benefit the community.
Acknowledgments
We extend our gratitude to our partners and collaborators across Southeast Asia.
Google’s Cloud AI and Research teams shared deep technical expertise and infrastructure support: enriching the dataset for Southeast Asian languages and leveraging Google Cloud’s Vertex Model Development Service for enhanced training efficiency and model development. This partnership builds on our existing collaboration between AI Singapore and Google on Project SEALD and the Aquarium open-source platform.
NCS for their partnership through the Pinnacle AI Industry Program pilot. Their Data & AI Engineers provided essential support for the development of our version 4 model, and in the process, gained valuable experience in LLM engineering. NCS is also the first deployment of version 4 in their Conversational AI Assistant, NCSgpt, accessible to over 10,000 NCS people in Asia-Pacific.
Nvidia’s engineering and technical resources for their support in developing version 4. Their assistance was crucial in enabling us to CPT Gemma on the NVIDIA Nemov2 Framework.
AI Singapore is a national programme supported by the National Research Foundation, Singapore and hosted by the National University of Singapore. Any opinion, finding, conclusion or recommendation expressed in this material are those of the author(s) and do not reflect the views of National Research Foundation, Singapore, or the National University of Singapore.
We also grateful for the support of the Infocomm Media Development Authority (IMDA) of Singapore.