SEA-LION v3

SEA-LION (Southeast Asian Languages In One Network) is AI Singapore's family of open LLMs continued-pretrained for the SEA region. Version 3 continued pretraining on ~200B tokens spanning 11 languages — Burmese, Chinese, English, Filipino, Indonesian, Khmer, Lao, Malay, Tamil, Thai, Vietnamese — on Llama-3.1 and Gemma-2 bases (8B, 9B, and 70B sizes), followed by instruction tuning on ~1.5M SEA instruction pairs. A v3.5 reasoning variant (8B-R) added hybrid reasoning/general modes.

It is the reference regional model: GoTo built its Sahabat-AI ecosystem on it, Qualcomm put an optimized build on the AI Hub for Snapdragon on-device deployment, and it anchors Singapore's National Multimodal LLM Programme. Models are on Hugging Face under Llama/Gemma community licenses with GGUF quants available.

Honest limits: as continued-pretraining derivatives, they inherit base-model weaknesses; low-resource SEA languages still underperform high-resource ones; reasoning depth lags frontier reasoning models; licensing follows the base model (Llama/Gemma community terms, not fully permissive).