Llama 3.3 70B Instruct: benchmark results
Meta Llama 3.3 70B instruction-tuned checkpoint. Provider: Meta. Released 2024-12-06. Access: Open.
Aggregate Rating 1573 ± 3, rank #534 of 1656 rated models, from 2880 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| ATAD - Bridge Sentence Evaluation | 60 | Accuracy (%) on the T4 bridge sentence evaluation task (pick | 100 |
| AbstentionBench - underspecified context - BB/Disambiguate - Recall | 97.33 | Abstention Recall (%) | 100 |
| AbstentionBench - underspecified context - Musique - F1 | 83.08 | Abstention F1 (%) | 100 |
| ArgBench (Chain-of-Thought) - Quality Assessment | 43.9 | Macro F1 (0-1, times 100) averaged over the argument quality | 100 |
| AspectSim (Sentence-Level Retrieval) | 0.59 | Spearman correlation (-1 to 1) between the embedding similar | 100 |
| BehaviorBench - Age from Personality | 9.3 | Mean absolute error in years of the age of the subject predi | 100 |
| BlueBench - Product Help | 86.64 | Score (%) | 100 |
| BlueBench - Summarization | 19.47 | Score (%) | 100 |
| CAT — MASK factual accuracy — Base | 78.8 | Accuracy (%) | 100 |
| CAT — MASK factual accuracy — Honest persona | 80 | Accuracy (%) | 100 |
| CAT — MASK factual accuracy — Honest persona + prefill | 78.9 | Accuracy (%) | 100 |
| CAT — MASK factual accuracy — Honest system | 77.8 | Accuracy (%) | 100 |
Interactive version: theaggregate.ai/model?slug=llama-3-3-70b-instruct · How It Works · Data refreshed daily, snapshot 2026-10-20.