Inference
EN
Microsoft

Phi-4-reasoning

Phi-4-reasoning is a reasoning model from Microsoft. 11 benchmarks count toward its index, in 6 categories.

  • Open weights
  • Reasoning
  • Input: text
  • Released 2025-04
RankedEnough results for an index.
IndexAll categories, out of 100.27.1±6.4Median 51.0
CoverageIndex weight that has results.85%
SpeedOutput tokens per second.41/sMedian 67/s
First tokenSeconds to the first output token.2.6 sMedian 1.3 s
Input / 1MUS dollars per 1M input tokens.—Median $0.60
Output / 1MUS dollars per 1M output tokens.—Median $2.55
ContextMaximum tokens in one request.32KMedian 262K
EloLMArena rating and rank.—Median 1422

Run locally

14.66B parameters
16-bit
31.8 GB
8-bit
17.4 GB
4-bit
10.3 GB
Weights
Hugging Face (opens in a new tab) ↗

Memory need with a short context. The 4-bit copy fits in 12 GB. Compare local models

Capabilities

Index per category, out of 100.
Out of 100
AgenticMulti-step tasks with tools.1 counted5.0—
CodingCode writing and repair.1 counted34.0—
ReasoningLogic problems and puzzles.2 counted22.1222 / 227
MultimodalTasks with images and text.None counted——
KnowledgeFacts and expert knowledge.3 counted33.3232 / 268
MultilingualTasks in many languages.None counted——
InstructionTasks with strict prompt rules.2 counted21.160 / 63
MathMath problems.2 counted35.5136 / 168

Results

11 counted
BenchmarkThe benchmark name.ResultThe score from the source.PlacePlace among results on this benchmark.
American Invitational Mathematics Examination 2024Math75.3%16 / 27
Instruction-Following EvalInstruction83.4%43 / 70
Massive Multitask Language Understanding ProfessionalKnowledge74.3%51 / 75
Critical Physics TasksReasoning0.0%150 / 214
Graduate-Level Google-Proof Q&AKnowledge65.8%101 / 141
American Invitational Mathematics Examination 2025Math62.9%39 / 53
LiveCodeBench: Holistic and Contamination Free Evaluation of Large Language Models for CodeCoding53.8%9 / 12
Artificial Analysis GPQA DiamondKnowledge57.5%163 / 199
Artificial Analysis Humanity's Last ExamKnowledge3.8%202 / 222
Artificial Analysis Long Context ReasoningReasoning0.0%204 / 217
Artificial Analysis IFBenchInstruction23.5%142 / 146
τ²-Bench Tool-Agent-User EvaluationAgentic0.0%148 / 149

Sources

BenchLMCC BY-NC 4.0

History

5 releases
Index27.6 → 27.1
23 Oct 202612 Nov 2026

Questions

How good is Phi-4-reasoning?
Its index is 27.1 out of 100. That is place 260 of 283 ranked models.
What is Phi-4-reasoning best at?
Math: 35.5 out of 100, place 136 of 168.
How fast is Phi-4-reasoning?
41 output tokens per second, 2.6 s to the first token.
What is the context window of Phi-4-reasoning?
32K tokens.
Is Phi-4-reasoning open source?
Yes. Its weights are published.
When was Phi-4-reasoning released?
April 2025.

Badge

Phi-4-reasoning on Inference360: place 260 of 283, index 27.1
HTML
<a href="https://inference360.ai/models/phi-4-reasoning"><img src="https://inference360.ai/badges/phi-4-reasoning.svg" alt="Phi-4-reasoning on Inference360: place 260 of 283, index 27.1" height="20"></a>
Markdown
[![Phi-4-reasoning on Inference360: place 260 of 283, index 27.1](https://inference360.ai/badges/phi-4-reasoning.svg)](https://inference360.ai/models/phi-4-reasoning)

Compare with

Gemini 2.5 Flash-Lite27.1Mistral Medium 327.1Llama 3.1 405B27.0Llama 4 Scout26.8Phi-422.4

More from Microsoft

All models
Phi-4-reasoning-plusUnrankedPhi-4 Multimodal InstructUnrankedPhi-422.4Phi-3.5-miniUnrankedWizardLM-2 8x22BUnranked