SEA-LION v4

model
AI Singapore

SEA-LION's first multimodal generation. Gemma-SEA-LION-v4-27B-IT continues pretraining of Gemma 3 27B on 500B tokens across Burmese, English, Mandarin, Indonesian, Khmer, Lao, Malay, Tagalog, Tamil, Thai, and Vietnamese, keeps Gemma's image input and 128K context, adds function calling, and at release ranked fifth of 55 models on SEA-HELM and first among open models under 200B parameters. Two Qwen-based siblings followed: Qwen-SEA-LION-v4-32B-IT on Qwen3-32B for reasoning (October 2025) and Qwen-SEA-LION-v4-8B-VL and 4B-VL on Qwen3-VL (November and December 2025), the 8B-VL being the org's most downloaded model. Released under the Gemma terms and MIT respectively. Not pretrained from scratch.

AnnouncementHuggingFace (Gemma-SEA-LION-v4-27B-IT)HuggingFace (Qwen-SEA-LION-v4-8B-VL)

Model Details

Architecture DENSE
Parameters 27B
Context window 131,072
Training tokens 500B
License Gemma Terms of Use
Base model gemma

Variants

Name Parameters Notes
Gemma-SEA-LION-v4-27B-IT 27B CPT of Gemma 3 27B, 500B tokens; image input; SEA-HELM #5 of 55 at release.
Qwen-SEA-LION-v4-32B-IT 32B Post-trained from Qwen3-32B for reasoning (October 2025), MIT.
Qwen-SEA-LION-v4-8B-VL 8B From Qwen3-VL-8B-Instruct (November 2025); a 4B-VL followed in December.
multilingualmultimodalvisionopen-weight

Related