AI Singapore releases Gemma-SEA-LION-Guard v2, MIT-licensed safety models for Southeast Asian languages
A 31B decision model (plus E4B and E2B) for content moderation and action governance across eight languages. Scores AUPRC 93.5 on general safety.
AI Singapore's AI Products Pillar, funded by the National Research Foundation Singapore, published Gemma-SEA-LION-Guard-v2 in October 2026. The 31B repository was created on 11 October, with E2B and E4B variants following; the card page shows "Last updated: 2026-10-21" and no release date. The 31B card calls it a safety decision model that outputs probability scores rather than free-form text, for content moderation and action governance.
It is built on Gemma-4-31B-IT, has 31B parameters in BF16, a 256K-token context and an MIT licence. Languages are Burmese, English, Indonesian, Filipino, Malay, Tamil, Thai and Vietnamese. Capabilities include multi-turn evaluation, bring-your-own-policy and tool-interaction evaluation.
On SEA-SafeguardBench the card reports AUPRC of 93.5 for general safety across English and Southeast Asian languages, and 78.7 for SEA-specific safety. Per-language scores range from 91.7 to 94.3.
The card warns that the model can hallucinate or produce ungrounded score distributions, especially near decision thresholds. It recommends human oversight and secondary verification before treating outputs as final. The authors disclaim liability for use of the released weights and code.
Public money, an open licence and a regional benchmark. The lower SEA-specific score is the number to watch.
Sources
Editor's first pass (AI-assisted)
- Real 4: MIT weights on the Hub; 31B, E4B and E2B variants.
- Useful 3.5: Covers Burmese, Indonesian, Filipino, Malay, Tamil, Thai and Vietnamese plus English; tool-interaction evaluation.
- Notable 3.5: A publicly funded regional safety model for eight languages.
- Novel 3.5: A regional benchmark and bring-your-own-policy support.
Comments