stablelm-zephyr-3B: benchmark results
Stability AI's StableLM Zephyr 3B, an instruction-tuned fine-tune of stablelm-3b-4e1t using DPO in a pipeline inspired by Hugging Face's Zephyr 7B, trained on UltraChat and MetaMathQA data. Provider: Stability AI. Released 2023-11-21. Access: Open.
Aggregate Rating 1317 ± 24, rank #1446 of 1656 rated models, from 38 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Open LLM Leaderboard - MuSR | 41.83 | Score | 58.4 |
| Open LLM Leaderboard v1 - GSM8K | 42.15 | Accuracy (%) (5-shot) | 56.9 |
| Open Portuguese LLM - FaQuAD NLI | 67.1 | Macro F1 (%) | 53.4 |
| Structured Local-Deployment MCQ | 54.19 | Strict accuracy (%): share of the 1,085 items answered with | 50 |
| RewardBench Chat Hard | 60.09 | Accuracy (%) | 46 |
| EuroEval Portuguese NLU - ScaLA PT | 8.98 | Matthews correlation coefficient (%) | 43.2 |
| EuroEval Portuguese NLU - MultiWikiQA PT | 65.92 | Character F1 (%) | 41.5 |
| Open Portuguese LLM - ASSIN2 RTE | 87.57 | Macro F1 (%) | 41.2 |
| EuroEval Portuguese NLU - SST-2 PT | 72.76 | Matthews correlation coefficient (%) | 40.7 |
| EuroEval Portuguese NLU | 45.29 | NLU Average Score (%) | 39.4 |
| RewardBench Safety | 74.05 | Accuracy (%) | 38.6 |
| RewardBench Reasoning | 75.73 | Accuracy (%) | 37.9 |
Interactive version: theaggregate.ai/model?slug=stablelm-zephyr-3b · How It Works · Data refreshed daily, snapshot 2026-10-20.