| Base | Localized Model |
|---|---|
| SEA-LION v1 | → WangchanLion 7B (Thai): WangchanLion 7B is a multilingual instruction-following model developed by PyThaiNLP and the VISTEC-depa AI Research Institute of Thailand. Fine-tuned on SEA-LION-v1-7B, it incorporates approximately 500,000 samples from open-source, commercially permissible datasets, with a focus on Thai and English languages. |
| SEA-LION v2 | → Llama3 8B CPT Sahabat-AI v1 Instruct (Indonesian): Llama3 8B CPT Sahabat-AI v1 Instruct model, co-developed by GoTo Group and AI Singapore, is an Indonesian-focused adaptation fine-tuned with 448,000 Indonesian instruction-completion pairs, along with 96,000 Javanese and 98,000 Sundanese pairs. It supports Indonesian, Javanese, Sundanese, and English, making it a significant advancement in AI for the Indonesian linguistic landscape. |
| SEA-LION v3 | → Gemma2 9B WangchanLIONv2 (Thai): The Gemma2 9B WangchanLIONv2 Instruct model is a collaborative effort between VISTEC and AI Singapore. It has been fine-tuned with approximately 3,760,000 Thai instruction-completion pairs derived from human-annotated instructions, FLAN-style automatic data construction, and synthetic samples. This multilingual model supports both Thai and English languages. → Gemma2 9B CPT Sahabat-AI (Indonesian): The Gemma2 9B CPT Sahabat-AI v1 Instruct model, co-developed by GoTo Group and AI Singapore, has been fine-tuned with approximately 448,000 Indonesian instruction-completion pairs, along with 96,000 in Javanese, 98,000 in Sundanese, and an additional 129,000 in English. This multilingual model supports Indonesian, Javanese, Sundanese, and English. |
| SEA-LION v4 | → Gemma 3 27B/4B SEA-LION v4 (Multimodal): The first multimodal release in the SEA-LION family, based on the Gemma 3 architecture. Co-developed with Google, these models support a 128K context window. It underwent post-training on ~10M samples across 11 SEA languages and features vision-text capabilities for document understanding and visual Q&A. This model supports text and image understanding with a commercially permissive license. It is designed to handle Southeast Asian cultural nuances and visual contexts. → Qwen-SEA-LION-v4-4B/8B-IT: The lightweight flagship multimodal model built on the Qwen3-VL framework, optimized for regional linguistic and cultural nuances. → Apertus 8B SEA-LION v4: A fully open regional adaptation based on the Swiss AI "Apertus" architecture, focusing on highly efficient, transparent, and community-driven multilingual support. Features 15T token pre-training across 1,000+ languages with a focus on deep multilingual depth and transparency. → Gemma-SEA-LION-v4.5-E2B-IT: Highly efficient lightweight model built on the Gemma 4 architecture, optimized for precise function-calling, structured JSON outputs, and autonomous agentic tool use in resource-constrained environments. → Qwen-SEA-LION-v4.5-27B-IT: High-capacity flagship instruction-tuned and multimodal model built on Qwen 3.6. Features a native 262K context window, robust multi-turn reasoning, and deep regional fluency across 7 key SEA languages. High-throughput with speculative decoding booster model utilizing block diffusion algorithms to significantly accelerate production inference latency without sacrificing output quality. → Nemotron-SEA-LION-v4.8: Extended regional foundation model featuring broad coverage across 13 regional languages—including Burmese, Malay, and key sub-national languages like Balinese, Javanese, and Sundanese. Advanced continue pre-training on parallel corpus to post-training methodologies—such as agentic on-policy distillation delivers state-of-the-art regional fluency and high computational efficiency, cultural competence assessments and country-entity knowledge verification. |
| SEA-Guard (renamed as SEA-LION-GUARD) | A specialized collection of safety-focused LLMs built on the SEA-LION family released on 4 Feb 2026. → Qwen-SEA-Guard-4B/8B (Image-to-Text): The 4B variant is a lightweight visual guardrail optimized for edge applications, while the 8B offers balanced visual moderation with stronger reasoning capabilities. → Llama-SEA-Guard-8B (Text Generation): A text-only safety specialist optimized for chat moderation and policy enforcement, ensuring safe interactions in dialogue systems. → Gemma-SEA-Guard-12B (Multimodal): The high-capacity safety flagship capable of interpreting complex relationships between visual and textual data for deep content analysis. Next-generation safety and action governance models featuring Bring Your Own Policy (BYOP) enforcement, real-time tool interception, and multi-turn context evaluation across 8 SEA languages. → Gemma-SEA-LION-Guard-v2-E2B (128K Context): An ultra-lightweight edge guardrail optimized for low-latency tool interaction evaluation and rapid API checks on constrained devices. → Gemma-SEA-LION-Guard-v2-E4B (128K Context): An efficient, mid-tier guardrail delivering balanced reasoning speed for real-time action governance in local enterprise deployments. → Gemma-SEA-LION-Guard-v2-31B (256K Context): The high-capacity flagship model built on Gemma-4, designed for deep policy evaluation, long-context scam detection, and multi-turn jailbreak analysis. |
| SEA-LION Embedding | A suite of high-performance encoder-only models specifically architected for Southeast Asian languages, providing state-of-the-art vector representations for RAG and semantic search. → SEA-LION-ModernBERT-Embedding (Encoder): Our flagship encoder line utilizing the ModernBERT architecture. It offers superior efficiency and long-context handling, specifically tuned for the unique scripts and linguistic structures of the SEA region. → SEA-LION-E5-Embedding (Encoder): A high-precision semantic encoder line fine-tuned from the E5-large foundation, optimized for maximum accuracy in cross-lingual retrieval and semantic matching across regional and global languages. |