SEA-LION: An Open-Source Large Language Model Family Built Exclusively for Southeast Asia

SEA-LION, an open-source project led by AI Singapore, is a family of large language models built exclusively for the diverse languages, cultures, and contexts of Southeast Asia. It includes multiple versions ranging from 3B to 70B parameters, supporting both text and multimodal tasks. Its core mission is to narrow the regional technological gap and ensure the benefits of AI are equitably shared among Southeast Asian users.

Southeast Asia has a population of over 670 million, thousands of languages and dialects, and rich cultural diversity. However, mainstream large models have weak support for local languages, leading to technological gaps such as understanding biases and cultural misinterpretations. The SEA-LION project emerged to address this, led by AI Singapore, aiming to provide better AI support for low-resource languages and underrepresented groups in Southeast Asia, and promote equitable access to technology.

SEA-LION has undergone multiple version iterations:

• v1: Pre-trained 3B/7B parameter models from scratch, laying the foundational architecture;
• v2: Continued pre-training based on Llama3, expanded context window to 8192 tokens, outperforming peers in Southeast Asian tasks;
• v3: Based on Gemma2/Llama3.1 architecture, launched 9B/8B/70B parameter models with context length up to 128K tokens, surpassing similar open-source models in both general and regional capabilities;
• v4: Introduced multimodal capabilities, supporting image+text input. Based on Gemma3/Qwen3-VL architecture, with 4B/8B/27B parameters, native context of 256K tokens, and optimized for Southeast Asian OCR scenarios;
• v4.5: Enhanced reasoning and tool usage capabilities through knowledge distillation and model merging.

Core training strategies: Continuous Pre-training (CPT) to inject knowledge from Southeast Asian corpora; Supervised Fine-tuning (SFT) to improve instruction following and dialogue quality. Architecture selection: Early self-pre-training, followed by adoption of mature open-weight models like Llama/Gemma/Qwen to balance performance and cost. Embedding models: Pre-trained 300M/600M parameter models from scratch based on the ModernBERT architecture, setting records on the SEA-BED benchmark, and providing leading performance in tasks like retrieval and re-ranking for 10 regional languages.

Released in February 2026, SEA-Guard is a safety companion model for SEA-LION, built on v4's multimodal capabilities and v3.5's reasoning abilities, providing a culturally adapted safety layer. Unlike general filters, it focuses on Southeast Asian cultural sensitivities and social norms, identifying and handling sensitive topics such as religion, politics, and ethnicity to ensure outputs align with local values.

Evaluation system includes traditional NLP benchmarks and the SEA-HELM diagnostic test (designed by regional language experts, covering four dimensions: English performance, Southeast Asian language proficiency, instruction following, and linguistic tasks). Performance: Each version consistently outperforms similar models. v1 surpassed most existing models upon release; v3 outperformed similar open-source models in both general and regional capabilities; v4's multimodal features opened new possibilities for complex scenarios.

Open-source ecosystem: Model weights and resources are released under the MIT license. Pre-training data, code, fine-tuning data, and evaluation benchmarks are publicly available. Community contributions are welcome (reporting issues, improving documentation, expanding language versions, etc.). Application scenarios: Government services (multilingual citizen consultation), education (local language personalized tutoring), business (reaching multicultural users). Model selection: From lightweight models for edge devices (3B/4B) to large models for cloud (27B/70B), supporting multiple quantization formats (GGUF, GPTQ, etc.) to lower deployment barriers.

SEA-LION demonstrates the value and feasibility of regionalized large models. Against the backdrop of dominant global models, it is expected to become the preferred foundational model for AI application development in Southeast Asia, providing solid technical support for the region's digital transformation.