Falcon3-3B-Instruct: benchmark results
TII's Falcon3-3B-Instruct is a 22-layer decoder-only Transformer instruction model using grouped-query attention and a publisher-supported 32K context. It is pruned and healed from Falcon3-7B-Base using 100 billion tokens. Instruction post-training uses 1.2 million samples. Provider: TII. Released 2024-12-17. Access: Open.
Aggregate Rating 1309 ± 19, rank #1438 of 1635 rated models, from 103 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Open LLM Leaderboard - IFEval | 69.77 | Score | 82.8 |
| MEDIC Benchmark | 75.28 | MEDIC Public Table Average | 82.1 |
| Open LLM Leaderboard - MATH Level 5 | 25 | Score | 78 |
| EuroEval Portuguese NLU - HAREM | 44.97 | Micro F1 excluding MISC (%) | 61.8 |
| Open Portuguese LLM - FaQuAD NLI | 70.01 | Macro F1 (%) | 60.2 |
| EuroEval Portuguese NLU - MultiWikiQA PT | 70.43 | Character F1 (%) | 59.4 |
| EuroEval Portuguese NLU - ScaLA PT | 14.58 | Matthews correlation coefficient (%) | 53.9 |
| Open LLM Leaderboard - MuSR | 41.36 | Score | 53.7 |
| EuroEval Portuguese NLU | 49.39 | NLU Average Score (%) | 49.2 |
| EuroEval Portuguese | 41.48 | Average Score (%) | 46.1 |
| Open LLM Leaderboard - GPQA | 28.86 | Score | 43.2 |
| Open Japanese LLM - Xlsum JA Rouge1 | 24.23 | Score (%) | 40.9 |
Interactive version: theaggregate.ai/model?slug=falcon3-3b-instruct · How It Works · Data refreshed daily, snapshot 2026-11-05.