Microsoft Research & Azureopen-weight
Released: 2025-02-15Phi-4-mini
Ultra-compact 3.8B parameter model featuring a 128K context window engineered for edge devices and high-volume routing.
Intelligence Index
75.2/ 100
Calibrated multi-domain compositeThroughput Speed
195tok/sec
Output streaming throughputTime To First Token
170ms
Initial chunk server latencyLiveBench Contam-Free
68.4%
September 2026 suite scoreSourced Benchmark Evaluations (61)
Verified performance across authoritative benchmarks with provenance tracking.| Benchmark | Status | Score | Trust | Date | Source Type | Provenance |
|---|---|---|---|---|---|---|
| ARC-AGI-3 | Active | 49.4% | 93 | 2025-03-22 | independent | Source ↗ |
| ARC-AGI-2 | Nearing Saturation | 59.3% | 75 | 2025-03-15 | independent | Source ↗ |
| FrontierMath | Active | 22% | 98 | 2025-03-12 | independent | Source ↗ |
| PutnamBench | Active | 40.8% | 90 | 2025-03-11 | independent | Source ↗ |
| SWE-Lancer | Deprecated | 48.3% | 20 | 2025-03-09 | independent | Source ↗ |
| Omni-MATH | Deprecated | 58.1% | 10 | 2025-03-09 | independent | Source ↗ |
| Soohak | Active | 59.7% | 98 | 2025-03-07 | independent | Source ↗ |
| FrontierCode | Active | 41.2% | 98 | 2025-03-07 | independent | Source ↗ |
| Humanity's Last Exam | Active | 33.4% | 93 | 2025-03-05 | independent | Source ↗ |
| StructFlowBench | Nearing Saturation | 58.8% | 75 | 2025-03-05 | independent | Source ↗ |
| AIME | Nearing Saturation | 74.6% | 75 | 2025-03-05 | independent | Source ↗ |
| OlympiadBench | Active | 64.4% | 80 | 2025-03-04 | independent | Source ↗ |
| MathArena | Active | 62.8% | 99 | 2025-03-03 | independent | Source ↗ |
| FireBench | Active | 58.1% | 95 | 2025-03-03 | independent | Source ↗ |
| Terminal-Bench | Active | 49.7% | 80 | 2025-03-03 | independent | Source ↗ |
| OpenAI-MRCR | Active | 63.2% | 80 | 2025-03-02 | independent | Source ↗ |
| SimpleQA | Active | 39.1% | 90 | 2025-03-02 | independent | Source ↗ |
| CodeContests | Deprecated | 46.2% | 10 | 2025-03-02 | independent | Source ↗ |
| ComplexBench | Nearing Saturation | 56.7% | 80 | 2025-03-01 | independent | Source ↗ |
| SWE-bench Pro | Active | 44% | 90 | 2025-03-01 | independent | Source ↗ |
| LongBench-Pro | Active | 61.7% | 80 | 2025-03-01 | independent | Source ↗ |
| ZeroSCROLLS | Nearing Saturation | 60.2% | 78 | 2025-02-28 | independent | Source ↗ |
| MultiPL-E | Deprecated | 65.3% | 10 | 2025-02-28 | independent | Source ↗ |
| LiveCodeBench | Saturated | 58.2% | 39 | 2025-02-27 | independent | Source ↗ |
| AgentIF | Nearing Saturation | 47% | 80 | 2025-02-27 | independent | Source ↗ |
| MATH | Saturated | 76.9% | 25 | 2025-02-27 | independent | Source ↗ |
| InfiniteBench | Nearing Saturation | 63.9% | 65 | 2025-02-27 | independent | Source ↗ |
| Mercury | Deprecated | 62.5% | 10 | 2025-02-26 | independent | Source ↗ |
| SCROLLS | Nearing Saturation | 60.2% | 65 | 2025-02-26 | independent | Source ↗ |
| MMLU-Pro | Nearing Saturation | 69.2% | 75 | 2025-02-26 | independent | Source ↗ |
| RULER | Nearing Saturation | 69.2% | 84 | 2025-02-25 | independent | Source ↗ |
| TabMWP | Saturated | 72.2% | 25 | 2025-02-25 | independent | Source ↗ |
| CyberSecEval | Deprecated | 63.9% | 10 | 2025-02-25 | independent | Source ↗ |
| SWE-bench Verified | Saturated | 55.4% | 25 | 2025-02-24 | independent | Source ↗ |
| MMLU-Redux | Saturated | 73.7% | 25 | 2025-02-24 | independent | Source ↗ |
| QuALITY | Nearing Saturation | 66.2% | 78 | 2025-02-24 | independent | Source ↗ |
| LongBench | Nearing Saturation | 64.7% | 65 | 2025-02-23 | independent | Source ↗ |
| SWE-bench | Deprecated | 48.3% | 10 | 2025-02-23 | independent | Source ↗ |
| MGSM | Nearing Saturation | 74.6% | 65 | 2025-02-23 | independent | Source ↗ |
| AGIEval | Saturated | 65.4% | 25 | 2025-02-22 | independent | Source ↗ |
| QASPER | Nearing Saturation | 62.4% | 65 | 2025-02-22 | independent | Source ↗ |
| TruthfulQA | Saturated | 66.2% | 25 | 2025-02-22 | independent | Source ↗ |
| GSM-Hard | Deprecated | 75.4% | 10 | 2025-02-21 | independent | Source ↗ |
| SuperGLUE | Saturated | 77.5% | 25 | 2025-02-21 | independent | Source ↗ |
| CodeXGLUE | Deprecated | 66.7% | 23 | 2025-02-20 | independent | Source ↗ |
| MT-Bench | Nearing Saturation | 7.1% | 75 | 2025-02-20 | independent | Source ↗ |
| BIG-Bench Hard | Saturated | 74.5% | 25 | 2025-02-20 | independent | Source ↗ |
| WinoGrande | Saturated | 79% | 28 | 2025-02-19 | independent | Source ↗ |
| MBPP+ | Nearing Saturation | 72.4% | 65 | 2025-02-19 | independent | Source ↗ |
| IFEval | Deprecated | 67.8% | 25 | 2025-02-19 | independent | Source ↗ |
| GPQA | Nearing Saturation | 66.2% | 78 | 2025-02-19 | independent | Source ↗ |
| Needle-in-a-Haystack | Saturated | 95% | 44 | 2025-02-18 | independent | Source ↗ |
| HumanEval+ | Saturated | 76.7% | 25 | 2025-02-18 | independent | Source ↗ |
| CommonsenseQA | Saturated | 79.8% | 28 | 2025-02-18 | independent | Source ↗ |
| MBPP | Saturated | 75.3% | 25 | 2025-02-17 | independent | Source ↗ |
| GSM8K | Saturated | 84.8% | 25 | 2025-02-17 | independent | Source ↗ |
| GLUE | Saturated | 82.1% | 28 | 2025-02-17 | independent | Source ↗ |
| MMLU | Saturated | 76.7% | 25 | 2025-02-16 | independent | Source ↗ |
| HumanEval | Saturated | 80.9% | 25 | 2025-02-16 | independent | Source ↗ |
| ARC-AGI | Saturated | 71.4% | 25 | 2025-02-16 | independent | Source ↗ |
| LiveBench (Continuous) | Active | 68.4% | 94 | 2025-02-01 | independent | Source ↗ |
Local Hardware Execution (Ollama)Hardware Compatible
# 1. Pull and execute model locally with Ollama
ollama run phi-4-mini
# 2. Or invoke via local OpenAI-compatible endpoint
curl http://localhost:11434/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{
"model": "phi-4-mini",
"messages": [{"role": "user", "content": "Analyze reasoning chains on AIME 2026"}]
}'Evaluation & Sourcing Notes
Scores listed for Phi-4-mini represent verified evaluations extracted from official research papers, independent evaluation suites (HELM, LMSYS, OpenCompass, LiveBench), and verified audit reports.
All benchmarks marked as saturated or deprecated reflect historical performance where the benchmark no longer provides active discriminative power.