AttaQ — IBM's published adversarial-question safety/red-teaming benchmark
| # | Model | Vendor | Best score | Runs | Last seen |
|---|---|---|---|---|---|
| 1 | Granite 3.3 8B Instruct | IBM | 88.5 | 2 | 2026-08-23 |
| 2 | Granite 3.3 2B Instruct | IBM | 87.5 | 1 | 2026-06-11 |
| 3 | Granite 4.1 30B | IBM | 85.8 | 1 | 2026-06-15 |
| 4 | Granite 3.1 8B Instruct | IBM | 85.7 | 1 | 2026-06-11 |
| 5 | Granite 3.2 8B Instruct | IBM | 84.7 | 1 | 2026-06-11 |
| 6 | Granite 3.1 2B Instruct | IBM | 84.7 | 1 | 2026-06-11 |
| 7 | Llama 3.1 8B Instruct | Meta | 83.4 | 1 | 2026-06-11 |
| 8 | Granite 3.2 2B Instruct | IBM | 83.2 | 1 | 2026-06-11 |
| 9 | Granite 4.1 8B | IBM | 81.2 | 1 | 2026-06-15 |
| 10 | DeepSeek-R1-Distill-Llama-8B | DeepSeek | 42.9 | 1 | 2026-06-11 |
| 11 | DeepSeek-R1-Distill-Qwen-7B | DeepSeek | 42.5 | 1 | 2026-06-11 |
Best tracked score per model (default configuration; source-attributed and verification-tiered). Open a model for its full benchmark surface, provenance and pricing.