FACTS Grounding: leaderboard
Google DeepMind and Google Research benchmark for long-form factuality and grounding against provided document context up to 32k tokens.
Metric: Combined Score (%). Source: huggingface.co. 34 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek R1 Distill Qwen 14B | 45.76 | #844 |
| 2 | Llama 3.3 70B Instruct | 42.55 | #534 |
| 3 | Qwen 3 4B | 42.55 | #829 |
| 4 | Qwen 3 30B A3B | 42.55 | #493 |
| 5 | Qwen 3 32B | 41.7 | #443 |
| 6 | DeepSeek R1 0528 Qwen3 8B | 41.1 | #764 |
| 7 | DeepSeek R1 Distill Llama 8B | 40.68 | #1291 |
| 8 | Qwen 3 8B | 40 | #674 |
| 9 | Qwen 3 14B | 38.3 | #537 |
| 10 | Gemma 3 27B (IT) | 37.8 | #523 |
| 11 | MedGemma-4B-IT | 37.71 | #853 |
| 12 | Qwen 2.5 VL 32B Instruct | 35.74 | #453 |
| 13 | Llama 3.1 70B Instruct | 33.47 | #561 |
| 14 | Gemma 3 12B (IT) | 31.3 | #664 |
| 15 | Gemma 3 4B (IT) | 30 | #992 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=facts-grounding · How It Works · Data refreshed daily, snapshot 2026-10-20.