Falcon3-7B-Instruct: benchmark results
TII's Falcon3-7B-Instruct is a 28-layer decoder-only Transformer instruction model using grouped-query attention and a publisher-supported 32K context. Its pretraining uses 14 trillion tokens. Instruction post-training uses 1.2 million samples. Provider: TII. Released 2024-12-17. Access: Open.
Aggregate Rating 1376 ± 15, rank #1300 of 1654 rated models, from 250 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| EuroEval Portuguese NLU - MultiWikiQA PT | 78.3 | Character F1 (%) | 96.8 |
| Open LLM Leaderboard - MuSR | 48.27 | Score | 95.5 |
| Open LLM Leaderboard - IFEval | 76.12 | Score | 91.7 |
| Open LLM Leaderboard - MATH Level 5 | 40.86 | Score | 90.7 |
| Open PL LLM - PSC (multiple choice, 5-shot) | 87.5 | Binary F1 (%) | 87.1 |
| MEDIC Benchmark | 79.83 | MEDIC Public Table Average | 86.3 |
| Open Portuguese LLM - ASSIN2 RTE | 93.09 | Macro F1 (%) | 86.3 |
| French LLM Leaderboard - IFEval FR | 62.33 | Score (%) | 83.3 |
| Open Portuguese LLM - BLUEX | 61.34 | Accuracy (%) | 80.2 |
| EuroEval Portuguese NLU - SST-2 PT | 82.25 | Matthews correlation coefficient (%) | 78.6 |
| Open LLM Leaderboard - BBH | 56.32 | Score | 77.4 |
| Open Japanese LLM - Xlsum JA Rouge1 | 29.22 | Score (%) | 76.7 |
Interactive version: theaggregate.ai/model?slug=falcon3-7b-instruct · How It Works · Data refreshed daily, snapshot 2026-10-24.