Granite-4.1-3B
IBM · released 2026-04-01 · Apache-2.0 license
Granite-4.1-3B is IBM's 3B dense Apache-2.0 instruct model with strong instruction-following (IFEval 82.3) and tool calling (BFCL 60.8) for its size, plus MMLU-Pro 49.8.
Key specs
| Type | Local open-weight |
|---|---|
| Parameters | 3.4B total |
| Architecture | granite |
| Context window | 131K tokens |
| Knowledge cutoff | — |
| Modalities | text |
| Recommended backends | — |
| Minimum viable rig | — |
Benchmark scores
| GPQA Diamond | 31.7% |
|---|---|
| SWE-bench Verified | — |
| AIME | — |
| MMLU-Pro | 49.83% |
| BFCL v3 (tool use) | 60.8% |
| Composite score | 5.2 |
| Community rating | No reviews yet |
VRAM & disk per quantization
| Quant | VRAM | Disk | RAM | Context |
|---|---|---|---|---|
| Q4_K_M | 3.5 GB | 2 GB | — | 131K |
Strengths & weaknesses
Strengths: Strong IFEval (82.3) and tool calling (BFCL 60.8) for a 3B; Solid MMLU-Pro 49.8 at small scale; 131K context; Apache-2.0
Weaknesses: GPQA 31.7 limited at 3B; Sparse parametric factual knowledge