Llama 3.1 405B: benchmark results
Meta's largest open-weights Llama 3.1 model, a 405B dense frontier-class flagship with a 128K context (July 2024). Provider: Meta. Released 2024-07-23. Access: Open.
Aggregate Rating 1593 ± 8, rank #466 of 1656 rated models, from 242 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| CyberSecEval-3 Code Interpreter Abuse - Llama Guard 3 Output | 1 | Abuse compliance (%) | 100 |
| CyberSecEval-3 Cyberattack Compliance | 38 | Attack compliance (%) | 100 |
| CyberSecEval-3 Cyberattack Compliance - Llama Guard 3 Input and Output | 19 | Attack compliance (%) | 100 |
| CyberSecEval-3 Cyberattack Compliance - Llama Guard 3 Output | 26 | Attack compliance (%) | 100 |
| CyberSecEval-3 Software Vulnerability Exploitation | 49 | Software exploitation success (%) | 100 |
| EuroEval Dutch | 78.43 | Average Score (%) | 100 |
| EuroEval Dutch NLU | 80.98 | NLU Average Score (%) | 100 |
| EuroEval English | 79.72 | Average Score (%) | 100 |
| KOR-Bench - Base Models (3-shot) | 39.68 | Accuracy (%; three-shot, 1,250 questions) | 100 |
| EuroEval Norwegian NLU - Norquad | 79.66 | Character F1 (%) | 99.7 |
| EuroEval French NLU - FQuAD | 77.36 | Character F1 (%) | 99.5 |
| ARC Challenge (AI2) | 95.3 | Accuracy (%) | 99.3 |
Interactive version: theaggregate.ai/model?slug=llama-3-1-405b · How It Works · Data refreshed daily, snapshot 2026-10-20.