Phi 4 Mini Instruct
microsoft · released 2025-10-17 · mit license
Phi-4-mini-instruct is Microsoft's 3.8B open model (Feb 2025, MIT, 128K): MATH-500 71.8, MMLU-Pro 52.8, GPQA Diamond 36.9, excelling at math/reasoning for its size while limited on facts and hard tasks.
Key specs
| Type | Local open-weight |
|---|---|
| Parameters | 3.84B total |
| Architecture | phi3 |
| Context window | 131K tokens |
| Knowledge cutoff | 2024-06-01 |
| Modalities | text |
| Recommended backends | — |
| Minimum viable rig | — |
Benchmark scores
| GPQA Diamond | 36.9% |
|---|---|
| SWE-bench Verified | — |
| AIME | 10% |
| MMLU-Pro | 52.8% |
| BFCL v3 (tool use) | — |
| Composite score | 4.08 |
| Community rating | No reviews yet |
VRAM & disk per quantization
| Quant | VRAM | Disk | RAM | Context |
|---|---|---|---|---|
| Q4_K_M | 3.7 GB | 2.2 GB | — | 131K |
Strengths & weaknesses
Strengths: Strong math/reasoning for 3.8B (MATH-500 71.8); Good function-calling and instruction-following (IFEval 70.1); 128K context, MIT
Weaknesses: Limited factual knowledge typical of small models; Weak hardest reasoning (AIME 10.0) and competitive coding