SmolLM3 3B
HuggingFaceTB · released 2025-07-08 · apache-2.0 license
Hugging Face's SmolLM3-3B is a fully-open 3B dual-mode reasoner. No-thinking scores: IFEval 76.7, BFCL 92.3, GPQA Diamond 35.7, AIME 9.3; enabling extended thinking raises math sharply (AIME 36.7).
Key specs
| Type | Local open-weight |
|---|---|
| Parameters | 3.08B total |
| Architecture | smollm3 |
| Context window | 66K tokens |
| Knowledge cutoff | — |
| Modalities | text |
| Recommended backends | — |
| Minimum viable rig | — |
Benchmark scores
| GPQA Diamond | 35.7% |
|---|---|
| SWE-bench Verified | — |
| AIME | 9.3% |
| MMLU-Pro | — |
| BFCL v3 (tool use) | 92.3% |
| Composite score | 5.01 |
| Community rating | No reviews yet |
VRAM & disk per quantization
| Quant | VRAM | Disk | RAM | Context |
|---|---|---|---|---|
| Q4_K_M | 3.3 GB | 1.8 GB | — | 66K |
Strengths & weaknesses
Strengths: Best-in-class instruction following at 3B (IFEval 76.7); Excellent tool calling (BFCL 92.3); dual-mode reasoning, 128K; Fully open, Apache-2.0
Weaknesses: Weak coding in non-reasoning mode (LiveCodeBench 15.2); Low AIME without thinking (9.3); reasoning mode much higher