Llama 3.2 3B Instruct
Meta · released 2024-09-01 · Llama 3.2 Community License license
Llama 3.2 3B Instruct is Meta's on-device 3B (Dec-2023 cutoff): vendor IFEval 77.4, nearly matching the 8B on instruction-following, with MMLU 63.4 and MATH 48.0.
Key specs
| Type | Local open-weight |
|---|---|
| Parameters | 3.21B total |
| Architecture | llama |
| Context window | 131K tokens |
| Knowledge cutoff | 2023-12-01 |
| Modalities | text |
| Recommended backends | — |
| Minimum viable rig | — |
Benchmark scores
| GPQA Diamond | — |
|---|---|
| SWE-bench Verified | — |
| AIME | — |
| MMLU-Pro | — |
| BFCL v3 (tool use) | — |
| Composite score | 7.74 |
| Community rating | No reviews yet |
VRAM & disk per quantization
| Quant | VRAM | Disk | RAM | Context |
|---|---|---|---|---|
| Q4_K_M | 3.4 GB | 1.9 GB | — | 131K |
Strengths & weaknesses
Strengths: Exceptional instruction-following for size (IFEval 77.4); On-device friendly (~6.5GB), 128K context; Strong math for 3B (vendor MATH 48.0)
Weaknesses: Limited deep reasoning (vendor GPQA 32.8); Smaller knowledge base than 8B