Phi 3.5 Mini Instruct
microsoft · released 2024-08-16 · mit license
Phi-3.5-mini-instruct is Microsoft's 3.8B open model (Aug 2024, MIT, 128K) for reasoning/math/multilingual, with strong long-context retrieval (RULER 84.1) but only moderate instruction-following (IFEval 50.6).
Key specs
| Type | Local open-weight |
|---|---|
| Parameters | 3.82B total |
| Architecture | phi3 |
| Context window | 131K tokens |
| Knowledge cutoff | 2023-10-01 |
| Modalities | text |
| Recommended backends | — |
| Minimum viable rig | — |
Benchmark scores
| GPQA Diamond | — |
|---|---|
| SWE-bench Verified | — |
| AIME | — |
| MMLU-Pro | 47.4% |
| BFCL v3 (tool use) | — |
| Composite score | 5.02 |
| 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 reasoning/math for 3.8B (GSM8K 86.2); Solid long-context retrieval (RULER 84.1), 128K; Multilingual; MIT
Weaknesses: Moderate instruction-following (IFEval 50.6); Lower MMLU-Pro than later small models