SEA-LION v4 (Latest)

SEA-LION version 4, released in August 2025, represents our first collection of multimodal models trained on Southeast Asian text. Each model offers unique strengths tailored to specific needs:

Vision Language Models (VLMs)

• Gemma-SEA-LION-v4-4B-VL A lightweight model optimized for mobile and edge devices, bringing robust SEA language understanding to environments with strict memory and latency constraints.

• Gemma-SEA-LION-v4-27B-VL Our powerful vision-language model, expertly trained to interpret both images and text with a deep, nuanced understanding of Southeast Asian cultural contexts.

• Qwen-SEA-LION-v4-4B-VL and 8B-VL Our latest specialized vision-language models, featuring a native 256K context window and superior OCR capabilities for Indonesian, Thai, and Vietnamese. The 4B model is optimized for efficient edge deployment, while the 8B model is engineered for complex multi-modal reasoning.

Text Generation LLMs

• Apertus-SEA-LION-v4-8B A versatile, compact model designed for high throughput and general-purpose SEA language tasks, offering an optimal trade-off between size and capability for developers.

• Gemma-SEA-LION-v4-27B Suited for translation, abstractive summarisation, natural language inference, causal reasoning, metaphor understanding, question answering, paraphrasing, and sentiment analysis—applications where regional language support and advanced reasoning are critical.

• Gemma-SEA-LION-v4-27B-IT Suited for knowledge-intensive tasks and high-demand contexts where comprehensive language comprehension is essential.

  • Available Variants: Gemma-SEA-LION-v4-27B-IT-GGUF, NVFP4, FP8_Dynamic support inference on a range of consumer-grade GPUs and are compatible with various inference engines.

• Qwen-SEA-LION-v4-32B-IT Our flagship instruction-tuned model, designed for maximum performance in general SEA context tasks.

  • Available Variants: Qwen-SEA-LION-v4-32B-IT-4BIT / 8BIT] offer a near-perfect balance of performance and efficiency, making the power of SEA-LION accessible on resource-constrained hardware like consumer-grade GPUs.

SEA-LION v4 continues our mission to create language models that understand and respond with greater cultural awareness and depth across Southeast Asia. In addition, SEA-LION v4 has the ability to handle both image and text input as a multimodal model.

For detailed information of each of the SEA-LION v4 models, please refer to their individual documentation pages via the links above.