h2o-danube3-4B-base: benchmark results
H2O.ai's h2o-danube3-4b-base, a 4B foundation model using an adjusted Llama 2 architecture with the Mistral tokenizer, trained for an 8,192-token context and designed to run offline on phones. Provider: H2O.ai. Released 2024-07-04. Access: Open.
Aggregate Rating 1287 ± 39, rank #1506 of 1661 rated models, from 38 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Open PL LLM - CBD (generative, 5-shot) | 22.7 | Macro F1 (%) | 46.2 |
| Open LLM Leaderboard - GPQA | 29.11 | Score | 46 |
| Open Portuguese LLM - FaQuAD NLI | 54.33 | Macro F1 (%) | 36.5 |
| Open Portuguese LLM - OAB Exams | 37.54 | Accuracy (%) | 35.6 |
| Open Portuguese LLM - ENEM | 50.8 | Accuracy (%) | 34.7 |
| Open PL LLM - PolEmo2-IN (generative, 5-shot) | 69.81 | Accuracy (%) | 34.2 |
| Open PL LLM - KLEJ NER (generative, 5-shot) | 38.19 | Accuracy (%) | 33.3 |
| Open PL LLM - Generative | 31.61 | Average Generative Score (%) | 31.1 |
| Open Portuguese LLM - BLUEX | 36.44 | Accuracy (%) | 30.1 |
| Open PL LLM - PolEmo2-OUT (generative, 5-shot) | 61.54 | Accuracy (%) | 29.5 |
| Open LLM Leaderboard - MuSR | 37.78 | Score | 28.2 |
| Open PL LLM - PPC (multiple choice, 5-shot) | 55.4 | Accuracy (%) | 28.2 |
Interactive version: theaggregate.ai/model?slug=h2o-danube3-4b-base · How It Works · Data refreshed daily, snapshot 2026-10-23.