Olmo 3 Safety Evaluation - HarmBench: leaderboard
The model gets 320 HarmBench harmful prompts (standard, copyright and contextual) with no attack; the score is the share of responses WildGuard labels safe. Unlike HarmBench's own board, which ranks attacks against models, this is direct requests only.
Metric: Safe responses (%, 320 harmful prompts, WildGuard safety label, higher is better). Source: arxiv.org. 36 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | OLMo 3.1 32B Instruct | 96 | #2384 |
| 2 | OLMo 3 7B Instruct | 94.9 | #6860 |
| 3 | Olmo-3.1-32B-Instruct-DPO | 93.9 | #3174 |
| 4 | Gemma 2 27B (IT) | 90.8 | #1782 |
| 5 | Olmo-3.1-32B-Instruct-SFT | 90.5 | #3407 |
| 6 | OLMo-3.1 32B Think | 89.7 | #1443 |
| 7 | OLMo-2-0325-32B-Instruct | 89 | #2570 |
| 8 | OLMo-2-1124-7B-Instruct | 88.8 | #7343 |
| 9 | Qwen 2.5 32B Instruct | 87.9 | #1215 |
| 10 | Granite 3.3 8B Instruct | 86.3 | #3615 |
| 11 | Qwen 3 VL 8B Instruct | 84.6 | #981 |
| 12 | Apertus-70B-Instruct-2509 | 84.4 | #2639 |
| 13 | Qwen 3 VL 8B (Thinking) | 81.9 | #887 |
| 14 | Qwen 3 VL 32B Instruct | 80.6 | #585 |
| 15 | Qwen 2.5 7B Instruct | 80.1 | #2564 |
Interactive version: theaggregate.ai/benchmark?slug=olmo-3-safety-evaluation-harmbench · How It Works · Data refreshed daily, snapshot 2026-11-13.