Granite 4.0 H Small
ibm-granite · released 2025-09-16 · apache-2.0 license
Granite-4.0-H-Small is IBM's 32B (9B active) hybrid Mamba2-MoE instruct model (Oct 2025, Apache-2.0): MMLU-Pro 55.5, GPQA 40.6, IFEval 87.6, strong tool calling. Artificial Analysis index 23 (pre-release).
Key specs
| Type | Local open-weight |
|---|---|
| Parameters | 32.21B total · MoE, — active |
| Architecture | granitemoehybrid |
| Context window | 131K tokens |
| Knowledge cutoff | — |
| Modalities | text |
| Recommended backends | — |
| Minimum viable rig | — |
Benchmark scores
| GPQA Diamond | 40.63% |
|---|---|
| SWE-bench Verified | — |
| AIME | — |
| MMLU-Pro | 55.47% |
| BFCL v3 (tool use) | 64.69% |
| Composite score | 5.9 |
| Community rating | No reviews yet |
VRAM & disk per quantization
| Quant | VRAM | Disk | RAM | Context |
|---|---|---|---|---|
| Q4_K_M | 20.2 GB | 18.7 GB | — | 131K |
Strengths & weaknesses
Strengths: Strong instruction following (IFEval 87.6) and tool calling (BFCL 64.7); Efficient 32B hybrid Mamba2-MoE (9B active), 128K; Apache-2.0
Weaknesses: GPQA 40.6 trails frontier models; No reasoning mode in this release