ThaiSafetyBench — leaderboard
Thai language and Thai cultural-context safety benchmark testing model resistance to harmful-content prompts across risk categories.
Metric: Overall ASR (self-reported). Source: benchmarklist.com. 24 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Llama 3.2 1B | 37.66 |
| 2 | Llama 3.1 8B Instruct | 32.44 |
| 3 | Gemma 3 4B | 28.11 |
| 4 | Llama 3.2 3B | 26.08 |
| 5 | Llama 3.1 70B Instruct | 24.49 |
| 6 | Gemma 3 12B | 20.4 |
| 7 | Llama 3.3 70B Instruct | 16.87 |
| 8 | Qwen 2.5 7B Instruct | 16.09 |
| 9 | GPT-4o | 16.04 |
| 10 | Qwen 2.5 72B Instruct | 12.34 |
| 11 | Claude Sonnet 4.5 | 9.75 |
| 12 | GPT-5 | 4.43 |
Interactive version: theaggregate.ai/benchmark?slug=thaisafetybench · How the rankings work · Data refreshed daily, snapshot 2026-07-22.