Granite 4.0 H Small: benchmark results
IBM's Apache-2.0 workhorse from the Granite 4.0 line: a 32B hybrid Mamba-2/transformer MoE with 9B active, built for low-memory enterprise agents (October 2025). Provider: IBM. Released 2025-10-02. Access: Open.
Aggregate Rating 1534 ± 15, rank #665 of 1656 rated models, from 801 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| HELM AIR-Bench 2024 - #16.36: Derogatory comments | 100 | Refusal Rate (%) | 100 |
| EuroEval Bosnian | 57.76 | Average Score (%) | 96.7 |
| HELM AIR-Bench 2024 - #13.2: Threats | 100 | Refusal Rate (%) | 95.9 |
| MTRAG-UN (RAG) | 38 | algorithmic reference-based score RB_alg (MTRAG's combinatio | 95.8 |
| EuroEval Hungarian NLU - HuSST | 62.31 | Matthews correlation coefficient (%) | 95 |
| HELM AIR-Bench 2024 - #44.3: Sexual exploitation | 100 | Refusal Rate (%) | 94.2 |
| EuroEval Slovene NLU - MultiWikiQA SL | 70.68 | Character F1 (%) | 93.6 |
| HELM AIR-Bench 2024 - #14.25: Personality | 100 | Refusal Rate (%) | 92.4 |
| Enkrypt AI - Toxicity Risk | 0.36 | Risk Score | 92.3 |
| EuroEval Finnish NLU - Scandisent FI | 92.45 | Matthews correlation coefficient (%) | 91.8 |
| EuroEval French NLU - ScaLA FR | 58.09 | Matthews correlation coefficient (%) | 91.8 |
| HELM AIR-Bench 2024 - #17.2: Suggestive | 73.33 | Refusal Rate (%) | 91.3 |
Interactive version: theaggregate.ai/model?slug=granite-4-0-h-small · How It Works · Data refreshed daily, snapshot 2026-10-20.