SEA-LION v4: Multimodal, Efficient LLM for Southeast Asian Languages

SEA-LION v4 at a glance

AI Singapore (AISG), in collaboration with Google, has released SEA-LION v4, an open-source multimodal language model built on the Gemma 3 (27B) architecture. The model targets Southeast Asian languages, including low-resource languages and dialects, and combines text and image understanding with a commercially permissive license and easy deployability on common hardware.

Strong results on SEA-HELM

SEA-LION v4 was evaluated on the SEA-HELM benchmark, a multilingual suite designed specifically for Southeast Asian languages. Despite its relatively compact 27B parameter size, v4 ranks top among models under 200B parameters and places #5 out of 55 models overall. Highlights include:

Filipino: 74.53 (v4) vs. 74.09 (Gemma 3-27B)
Malay: 71.31 (v4) vs. 71.20 (Gemma 3-27B)
Tamil: 68.47 (v4) vs. 68.45 (Gemma 3-27B)
Burmese: 57.18 (v4), close to Gemma 3's 57.78 and outperforming larger MoE models

In several languages SEA-LION v4 matches or exceeds the performance of models that are 3–10x larger, demonstrating an effective balance of efficiency and capability for both research and production use.

Key technical improvements

SEA-LION v4 introduces several advances that make it suitable for regional and global applications:

Open sourcing and permissive licensing: released under the Gemma license for commercial use, with distribution across Hugging Face, Google Cloud Vertex AI, AWS SageMaker, Kaggle, NVIDIA NIM, and Ollama for edge deployment.
Efficiency and portability: quantized FP4 and FP8 versions deliver less than 0.5% performance drop compared to full precision, up to 50% faster inference, and the ability to run on consumer-grade hardware such as a 32GB RAM laptop.
Multimodality: first multimodal release for the SEA-LION initiative, supporting image understanding alongside text and offering 128K token context windows for extended reasoning over long documents and multi-turn prompts.
Agentic and structured interactions: built-in function calling, structured JSON/schema outputs, and compatibility with agentic workflows for real-world automation and orchestration.

Training focus and regional strengths

The model was trained on over 1 trillion tokens with a heavy emphasis on a curated Southeast Asian dataset. This specialized training improves performance on low-resource languages, dialects, and culturally specific contexts where many global models underperform. On SEA-HELM tasks for Filipino, Malay, Tamil, and Burmese, SEA-LION v4 consistently ranks among the top models across parameter ranges.

At the same time, inheriting Gemma's strong reasoning capabilities keeps SEA-LION v4 competitive on English and other global tasks, making it a versatile option for international deployment.

Deployment, use cases, and impact

SEA-LION v4 aims to lower barriers to high-quality multilingual and multimodal AI in the region. Typical use cases include:

Multilingual document analysis and translation with embedded images
Image-grounded question answering in local languages
Interactive agents and workflow automation with structured outputs

Because it is open and portable, the model supports both cloud-scale enterprise integrations and on-device scenarios, contributing to greater digital equity across Southeast Asia.

Where to find it

The model is available through Hugging Face and supported across multiple cloud and edge platforms. Resources such as tutorials, notebooks, and the SEA-LION playground are provided by the project for experimentation and adoption.