SEA-LION v4

SEA-LION v4 is the fourth generation of AI Singapore's Southeast Asian Languages In One Network flagship, released in late 2025. The v4 generation is a family rather than a single checkpoint: a Qwen3-32B-based variant released with Alibaba Cloud in November 2025, and a Gemma 3 27B-based multimodal variant trained with Google Cloud support.

Both were further trained on over 100 billion tokens of Southeast Asian language data to handle the region's low-resource languages, code-switching, and cultural context. The Qwen-based variant topped the SEA-HELM leaderboard among open models under 200B parameters and is sized to run on consumer hardware with 32GB of RAM.

In September 2026 the family expanded again with Nemotron-SEA-LION-v4.8, built with NVIDIA on Nemotron 3. The project is explicitly sovereign: it keeps Southeast Asian linguistic data inside a regional open model rather than as a footnote in US-centric training corpora.