Qwen2.5 14B Instruct
Qwen · released 2024-09-16 · apache-2.0 license
Qwen2.5-14B-Instruct is the cost-effective mid-tier Qwen2.5 model (Apache-2.0): vendor GPQA 45.5, MMLU-Pro 63.7, LiveCodeBench 42.6, positioned as competitive with GPT-4o-mini.
Key specs
| Type | Local open-weight |
|---|---|
| Parameters | 14.77B total |
| Architecture | qwen2 |
| Context window | 33K tokens |
| Knowledge cutoff | — |
| Modalities | text |
| Recommended backends | — |
| Minimum viable rig | — |
Benchmark scores
| GPQA Diamond | 45.5% |
|---|---|
| SWE-bench Verified | — |
| AIME | — |
| MMLU-Pro | 63.7% |
| BFCL v3 (tool use) | — |
| Composite score | 6.14 |
| Community rating | No reviews yet |
VRAM & disk per quantization
| Quant | VRAM | Disk | RAM | Context |
|---|---|---|---|---|
| Q4_K_M | 10.1 GB | 8.6 GB | — | 33K |
Strengths & weaknesses
Strengths: Positioned as competitive with GPT-4o-mini; Strong instruction following (IFEval 81.0); Apache-2.0
Weaknesses: No native reasoning mode; No vendor SWE-bench/AIME numbers