Benchmarks
347 sources · 1,258 models × 3,038 benchmarks · 158,912 scoresThe safety competency of SEA-HELM's Thai score: the mean of toxicity detection (classifying Thai text as clean, offensive or hateful) and the SEA-Safeguard task group (classifying content as safe or harmful under a 14-category safety policy, including culturally specific cases), scored by balanced accuracy rescaled against chance and floored at 0, from 8 runs per model with prompts in the language. Diagnostic: it measures safety classification rather than a model's own refusals, and reaches the SEA-HELM score only through the Thai score.
No signal or difficulty estimate. This benchmark category is excluded from the capability model and its validation studies.
| # | Model | Setting | Score | Source |
|---|---|---|---|---|
| 1 | Qwen3.6 27B | thinking · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); reasoning on (the board labels the row reasoning); effort or thinking budget unpublished | 61.3 | |
| 2 | Qwen3.5-27B | thinking · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); reasoning on (the board labels the row reasoning); effort or thinking budget unpublished | 60.5 | |
| 3 | Inkling-Small | thinking · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); reasoning on (the board labels the row reasoning); effort or thinking budget unpublished | 59.4 | |
| 4 | Qwen3.5 122B A10B | thinking · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); reasoning on (the board labels the row reasoning); effort or thinking budget unpublished | 59.0 | |
| 5 | Qwen3 14B | thinking · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); reasoning on (the board labels the row reasoning); effort or thinking budget unpublished | 58.6 | |
| 6 | Qwen3 VL 32B Instruct | not supported · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); the model has no reasoning mode | 58.4 | |
| 7 | Qwen3.8-Flash-Next | thinking · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); reasoning on (the board labels the row reasoning); effort or thinking budget unpublished | 58.3 | |
| 8 | Qwen3.5 397B A17B | thinking · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); reasoning on (the board labels the row reasoning); effort or thinking budget unpublished | 58.3 | |
| 9 | Qwen3.6 35B A3B | thinking · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); reasoning on (the board labels the row reasoning); effort or thinking budget unpublished | 57.9 | |
| 10 | Qwen3 32B | thinking · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); reasoning on (the board labels the row reasoning); effort or thinking budget unpublished | 57.8 | |
| 11 | Qwen3.5-9B | thinking · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); reasoning on (the board labels the row reasoning); effort or thinking budget unpublished | 57.5 | |
| 12 | Qwen3.5 35B A3B | thinking · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); reasoning on (the board labels the row reasoning); effort or thinking budget unpublished | 57.3 | |
| 13 | GPT-5.4 Mini | thinking · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); reasoning on (the board labels the row reasoning); effort or thinking budget unpublished | 56.7 | |
| 14 | Nemotron 3 Nano 30B A3B | thinking · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); reasoning on (the board labels the row reasoning); effort or thinking budget unpublished | 56.7 | |
| 15 | Command A+ (May 2026, BF16) | thinking · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); reasoning on (the board labels the row reasoning); effort or thinking budget unpublished | 56.4 | |
| 16 | GPT-5.5 | thinking · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); reasoning on (the board labels the row reasoning); effort or thinking budget unpublished | 56.0 | |
| 17 | MiniMax M3 | thinking · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); reasoning on (the board labels the row reasoning); effort or thinking budget unpublished | 55.5 | |
| 18 | Sonnet 4.6 | thinking · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); reasoning on (the board labels the row reasoning); effort or thinking budget unpublished | 55.3 | |
| 19 | Mistral Medium 3.5 | disabled · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); reasoning off (the board labels the row non-reasoning) | 55.1 | |
| 20 | Gemini 3.8 Flash | thinking · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); reasoning on (the board labels the row reasoning); effort or thinking budget unpublished | 54.9 | |
| 21 | gpt-oss-120b | thinking · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); reasoning on (the board labels the row reasoning); effort or thinking budget unpublished | 54.6 | |
| 22 | Qwen3.5-4B | thinking · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); reasoning on (the board labels the row reasoning); effort or thinking budget unpublished | 54.6 | |
| 23 | Qwen3 8B | thinking · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); reasoning on (the board labels the row reasoning); effort or thinking budget unpublished | 54.0 | |
| 24 | Nemotron 3 Ultra 550B A55B | disabled · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); reasoning off (the board labels the row non-reasoning) | 53.9 | |
| 25 | GPT-5.4 | thinking · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); reasoning on (the board labels the row reasoning); effort or thinking budget unpublished | 53.6 | |
| 26 | Gemma 4 26B A4B IT | thinking · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); reasoning on (the board labels the row reasoning); effort or thinking budget unpublished | 53.5 | |
| 27 | GLM-4.7 | thinking · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); reasoning on (the board labels the row reasoning); effort or thinking budget unpublished | 53.2 | |
| 28 | Gemma 4 31B IT | thinking · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); reasoning on (the board labels the row reasoning); effort or thinking budget unpublished | 52.8 | |
| 29 | Gemma 4 E4B IT | thinking · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); reasoning on (the board labels the row reasoning); effort or thinking budget unpublished | 52.7 | |
| 30 | Qwen3 VL 4B Instruct | not supported · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); the model has no reasoning mode | 52.4 | |
| 31 | Muse Glimmer 30B | thinking · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); reasoning on (the board labels the row reasoning); effort or thinking budget unpublished | 52.3 | |
| 32 | Sonnet 5 | thinking · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); reasoning on (the board labels the row reasoning); effort or thinking budget unpublished | 52.0 | |
| 33 | Gemini 3.1 Flash-Lite | thinking · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); reasoning on (the board labels the row reasoning); effort or thinking budget unpublished | 52.0 | |
| 34 | Gemini 3.1 Pro Preview | thinking · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); reasoning on (the board labels the row reasoning); effort or thinking budget unpublished | 51.5 | |
| 35 | Opus 4.8 | thinking · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); reasoning on (the board labels the row reasoning); effort or thinking budget unpublished | 51.3 | |
| 36 | Qwen3-VL-235B-A22B-Instruct | not supported · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); the model has no reasoning mode | 50.7 | |
| 37 | Haiku 4.5 | thinking · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); reasoning on (the board labels the row reasoning); effort or thinking budget unpublished | 50.7 | |
| 38 | gpt-oss-20b | thinking · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); reasoning on (the board labels the row reasoning); effort or thinking budget unpublished | 49.6 | |
| 39 | GPT-5.6 Sol | thinking · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); reasoning on (the board labels the row reasoning); effort or thinking budget unpublished | 49.1 | |
| 40 | Gemma 4 12B IT | disabled · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); reasoning off (the board labels the row non-reasoning) | 48.7 | |
| 41 | Qwen3-VL-8B-Instruct | not supported · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); the model has no reasoning mode | 48.6 | |
| 42 | GPT-6 Astra | thinking · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); reasoning on (the board labels the row reasoning); effort or thinking budget unpublished | 48.4 | |
| 43 | Llama 3.3 70B Instruct | not supported · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); the model has no reasoning mode | 48.1 | |
| 44 | GPT-5.6 Terra | thinking · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); reasoning on (the board labels the row reasoning); effort or thinking budget unpublished | 48.0 | |
| 45 | GPT-5.6 Luna | thinking · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); reasoning on (the board labels the row reasoning); effort or thinking budget unpublished | 47.5 | |
| 46 | Gemma 3 27B IT | not supported · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); the model has no reasoning mode | 46.9 | |
| 47 | Llama 4 Maverick Instruct | not supported · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); the model has no reasoning mode | 46.8 | |
| 48 | Gemini 3.6 Flash | thinking · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); reasoning on (the board labels the row reasoning); effort or thinking budget unpublished | 46.8 | |
| 49 | GLM 4.7 Flash | thinking · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); reasoning on (the board labels the row reasoning); effort or thinking budget unpublished | 46.7 | |
| 50 | Gemini 3.5 Flash | thinking · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); reasoning on (the board labels the row reasoning); effort or thinking budget unpublished | 46.0 | |
| 51 | Gemma 4 E2B IT | thinking · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); reasoning on (the board labels the row reasoning); effort or thinking budget unpublished | 45.2 | |
| 52 | Mistral Small 4 | disabled · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); reasoning off (the board labels the row non-reasoning) | 43.0 | |
| 53 | Gemma 3 12B IT | not supported · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); the model has no reasoning mode | 41.9 | |
| 54 | Llama 4 Scout Instruct | not supported · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); the model has no reasoning mode | 41.6 | |
| 55 | Llama 3.2 3B Instruct | not supported · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); the model has no reasoning mode | 37.2 | |
| 56 | Gemma 3 4B IT | not supported · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); the model has no reasoning mode | 35.4 | |
| 57 | Apertus 70B Instruct | not supported · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); the model has no reasoning mode | 33.0 | |
| 58 | Nemotron 3 Super 120B A12B | disabled · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); reasoning off (the board labels the row non-reasoning) | 32.3 | |
| 59 | Apertus 8B Instruct | not supported · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); the model has no reasoning mode | 30.8 | |
| 60 | Nemotron 3.5 Lightning 30B A3B BF16 | thinking · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); reasoning on (the board labels the row reasoning); effort or thinking budget unpublished | 30.4 | |
| 61 | Llama 3.1 8B Instruct | not supported · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); the model has no reasoning mode | 22.8 | |
| 62 | Olmo 3.1 32B Instruct | not supported · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); the model has no reasoning mode | 22.2 | |
| 63 | Tiny Aya Global | not supported · SEA-HELM leaderboard: zero-shot native-language prompts in seven Southeast Asian languages, 8 runs per model with the model's default generation settings (vLLM defaults otherwise); the model has no reasoning mode | 0.0 |
Benchmarks outside its family whose model rankings track it most reliably (Spearman ρ over at least 10 shared models).