Gemma-SEA-LION-Guard-v2-31B
Gemma-SEA-LION-Guard-v2-31B adapts Gemma 4 31B IT to score content and proposed actions. It can assess several conversation turns, apply user-supplied rules and inspect tool interactions. AI Singapore publishes weights and describes managed API access.
How this release works
The decision interface separates content moderation, rule applicability and rule violation. Content moderation scores safe, sensitive and unsafe categories. The model card lists Burmese, English, Indonesian, Filipino, Malay, Tamil, Thai and Vietnamese. These are decision outputs rather than a general chat interface.Gemma-SEA-LION-Guard-v2-31B official model card
How to assess it for your work
Practical guidance. Use this release where a decision depends on several earlier messages or detailed rules. Set the application's action threshold separately from the model score. Review borderline cases and cases where a rule does not apply. A high probability is a model judgment, not proof that the action meets a policy.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | API, Downloadable |
| Model identifiers | aisingapore/Gemma-SEA-LION-Guard-v2-31B |
| Base model | google/gemma-4-31B-itGemma-SEA-LION-Guard-v2-31B official model card |
| Context stated by the card | 256K tokensGemma-SEA-LION-Guard-v2-31B official model card |
| License declared by the model card | MITGemma-SEA-LION-Guard-v2-31B official model card |
| Output type | Structured probabilitiesGemma-SEA-LION-Guard-v2-31B official model card |
Related models
- Gemma-SEA-LION-Guard-v2-E2B
AI Singapore · SEA-LION Guard v2
A compact Gemma 4 E2B moderation model that evaluates conversations and tool calls against custom rules.
Launch date not confirmedAPI · Downloadable - Gemma-SEA-LION-Guard-v2-E4B
AI Singapore · SEA-LION Guard v2
A Gemma 4 E4B moderation model for custom policy checks across conversations and tool interactions.
Launch date not confirmedAPI · Downloadable