Olmo 3 Safety Evaluation: leaderboard
Each model answers the safety suite Ai2 used while training Olmo 3 (harmful requests, jailbreaks and benign look-alikes); the score is the paper's own average of six sets, with WildJailbreak counted as the mean of harmful refusal and benign compliance. Unlike HELM Safety, one publisher re-ran current open models (Olmo 3, Qwen 3, Gemma 3, Nemotron) with one judge.
Metric: Mean safety score of six development sets (%, higher is better). Source: arxiv.org. 36 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Olmo-3.1-32B-Instruct-SFT | 92.1 | #3407 |
| 2 | OLMo-2-1124-7B-Instruct | 91.1 | #7343 |
| 3 | OLMo 3.1 32B Instruct | 89.5 | #2384 |
| 4 | Olmo-3.1-32B-Instruct-DPO | 88.9 | #3174 |
| 5 | OLMo 3 7B Instruct | 87.6 | #6860 |
| 6 | Qwen 3 VL 32B Instruct | 85.8 | #585 |
| 7 | OLMo-2-0325-32B-Instruct | 84.2 | #2570 |
| 8 | OLMo-3.1 32B Think | 83.6 | #1443 |
| 9 | Qwen 3 VL 8B (Thinking) | 82.9 | #887 |
| 10 | Qwen 3 VL 32B (Thinking) | 82.7 | #559 |
| 11 | Qwen 2.5 32B Instruct | 82.2 | #1215 |
| 12 | Qwen 3 32B (Non-reasoning) | 81.6 | #960 (Qwen 3 32B) |
| 13 | Qwen 3 8B (Non-reasoning) | 78.4 | #1830 (Qwen 3 8B) |
| 14 | Qwen 3 VL 8B Instruct | 77.7 | #981 |
| 15 | Apertus-70B-Instruct-2509 | 77.1 | #2639 |
Interactive version: theaggregate.ai/benchmark?slug=olmo-3-safety-evaluation · How It Works · Data refreshed daily, snapshot 2026-11-13.