Microsoft Research & Azureopen-weight
Released: 2025-01-08Phi-4
Microsoft high-efficiency 14.7B parameter model matching earlier 70B models on math reasoning and programmatic logic.
Intelligence Index
84.8/ 100
Calibrated multi-domain compositeThroughput Speed
105tok/sec
Output streaming throughputTime To First Token
260ms
Initial chunk server latencyLiveBench Contam-Free
76.5%
September 2026 suite scoreSourced Benchmark Evaluations (51)
Verified performance across authoritative benchmarks with provenance tracking.| Benchmark | Status | Score | Trust | Date | Source Type | Provenance |
|---|---|---|---|---|---|---|
| ARC-AGI-3 | Active | 55.3% | 93 | 2025-02-12 | independent | Source ↗ |
| ARC-AGI-2 | Nearing Saturation | 66.3% | 75 | 2025-02-05 | independent | Source ↗ |
| FrontierMath | Active | 24.2% | 98 | 2025-02-02 | independent | Source ↗ |
| LiveBench (Continuous) | Active | 76.5% | 94 | 2025-02-01 | independent | Source ↗ |
| PutnamBench | Active | 44.9% | 90 | 2025-02-01 | independent | Source ↗ |
| SWE-Lancer | Deprecated | 56.1% | 20 | 2025-01-30 | independent | Source ↗ |
| Omni-MATH | Deprecated | 63.9% | 10 | 2025-01-30 | independent | Source ↗ |
| Soohak | Active | 65.7% | 98 | 2025-01-28 | independent | Source ↗ |
| FrontierCode | Active | 47.8% | 98 | 2025-01-28 | independent | Source ↗ |
| AIME | Nearing Saturation | 82.1% | 75 | 2025-01-26 | independent | Source ↗ |
| StructFlowBench | Nearing Saturation | 66.3% | 75 | 2025-01-26 | independent | Source ↗ |
| Humanity's Last Exam | Active | 37.4% | 93 | 2025-01-26 | independent | Source ↗ |
| OlympiadBench | Active | 70.8% | 80 | 2025-01-25 | independent | Source ↗ |
| MathArena | Active | 69.1% | 99 | 2025-01-24 | independent | Source ↗ |
| Terminal-Bench | Active | 57.7% | 80 | 2025-01-24 | independent | Source ↗ |
| FireBench | Active | 65.5% | 95 | 2025-01-24 | independent | Source ↗ |
| SimpleQA | Active | 44.1% | 90 | 2025-01-23 | independent | Source ↗ |
| CodeContests | Deprecated | 53.6% | 10 | 2025-01-23 | independent | Source ↗ |
| ComplexBench | Nearing Saturation | 64% | 80 | 2025-01-22 | independent | Source ↗ |
| SWE-bench Pro | Active | 51.2% | 90 | 2025-01-22 | independent | Source ↗ |
| MultiPL-E | Deprecated | 75.9% | 10 | 2025-01-21 | independent | Source ↗ |
| LiveCodeBench | Saturated | 67.6% | 39 | 2025-01-20 | independent | Source ↗ |
| MATH | Saturated | 84.7% | 25 | 2025-01-20 | independent | Source ↗ |
| AgentIF | Nearing Saturation | 53.1% | 80 | 2025-01-20 | independent | Source ↗ |
| Mercury | Deprecated | 72.6% | 10 | 2025-01-19 | independent | Source ↗ |
| MMLU-Pro | Nearing Saturation | 78% | 75 | 2025-01-19 | independent | Source ↗ |
| CyberSecEval | Deprecated | 74.3% | 10 | 2025-01-18 | independent | Source ↗ |
| TabMWP | Saturated | 79.5% | 25 | 2025-01-18 | independent | Source ↗ |
| SWE-bench Verified | Saturated | 64.4% | 25 | 2025-01-17 | independent | Source ↗ |
| MMLU-Redux | Saturated | 83.1% | 25 | 2025-01-17 | independent | Source ↗ |
| SWE-bench | Deprecated | 56.1% | 10 | 2025-01-16 | independent | Source ↗ |
| MGSM | Nearing Saturation | 82.1% | 65 | 2025-01-16 | independent | Source ↗ |
| TruthfulQA | Saturated | 74.6% | 25 | 2025-01-15 | independent | Source ↗ |
| AGIEval | Saturated | 73.1% | 25 | 2025-01-15 | independent | Source ↗ |
| SuperGLUE | Saturated | 86.7% | 25 | 2025-01-14 | independent | Source ↗ |
| GSM-Hard | Deprecated | 82.9% | 10 | 2025-01-14 | independent | Source ↗ |
| MT-Bench | Nearing Saturation | 8% | 75 | 2025-01-13 | independent | Source ↗ |
| BIG-Bench Hard | Saturated | 83.3% | 25 | 2025-01-13 | independent | Source ↗ |
| CodeXGLUE | Deprecated | 77.6% | 23 | 2025-01-13 | independent | Source ↗ |
| MBPP+ | Nearing Saturation | 84.2% | 65 | 2025-01-12 | independent | Source ↗ |
| IFEval | Deprecated | 76.5% | 25 | 2025-01-12 | independent | Source ↗ |
| WinoGrande | Saturated | 88.4% | 28 | 2025-01-12 | independent | Source ↗ |
| GPQA | Nearing Saturation | 74.6% | 78 | 2025-01-12 | independent | Source ↗ |
| CommonsenseQA | Saturated | 89.3% | 28 | 2025-01-11 | independent | Source ↗ |
| HumanEval+ | Saturated | 89.1% | 25 | 2025-01-11 | independent | Source ↗ |
| GLUE | Saturated | 91.8% | 28 | 2025-01-10 | independent | Source ↗ |
| MBPP | Saturated | 87.5% | 25 | 2025-01-10 | independent | Source ↗ |
| GSM8K | Saturated | 93.3% | 25 | 2025-01-10 | independent | Source ↗ |
| MMLU | Saturated | 86.5% | 25 | 2025-01-09 | independent | Source ↗ |
| HumanEval | Saturated | 94.1% | 25 | 2025-01-09 | independent | Source ↗ |
| ARC-AGI | Saturated | 79.9% | 25 | 2025-01-09 | independent | Source ↗ |
Local Hardware Execution (Ollama)Hardware Compatible
# 1. Pull and execute model locally with Ollama
ollama run phi-4
# 2. Or invoke via local OpenAI-compatible endpoint
curl http://localhost:11434/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{
"model": "phi-4",
"messages": [{"role": "user", "content": "Analyze reasoning chains on AIME 2026"}]
}'Evaluation & Sourcing Notes
Scores listed for Phi-4 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.