Phi-4

Unknown Developer

Phi-4 is a cutting-edge open-source model designed for superior performance in complex reasoning, coding, and knowledge-intensive applications. Its capabilities are powered by a carefully curated mix of synthetic data, refined web content, scholarly materials, and supervised fine-tuning, ensuring accuracy, ethical alignment, and safety.

Model Specifications

Technical details and capabilities of Phi-4

Core Specifications

14.7B Parameters

Model size and complexity

9800.0B Training Tokens

Amount of data used in training

16.0K / 16.0K

Input / Output tokens

May 31, 2024

Knowledge cutoff date

December 11, 2024

Release date

Capabilities & License

Multimodal Support

Not Supported

Web Hydrated

License

MIT

Resources

Research Paper

https://arxiv.org/pdf/2412.08905

API Reference

https://huggingface.co/microsoft/phi-4

Performance Insights

Check out how Phi-4 handles various AI tasks through comprehensive benchmark results.

84.8

MMLU

84.8

(94%)

82.8

HumanEval+

82.8

(92%)

82.6

HumanEval

82.6

(92%)

80.6

MGSM

80.6

(90%)

80.4

MATH

80.4

(89%)

75.5

DROP

75.5

(84%)

75.4

ArenaHard

75.4

(84%)

70.4

MMLU-Pro

70.4

(78%)

IFEval

(70%)

56.2

PhiBench

56.2

(62%)

56.1

GPQA

56.1

(62%)

47.6

LiveBench

47.6

(53%)

SimpleQA

(3%)

MMLU

HumanEval+

HumanEval

MGSM

MATH

DROP

ArenaHard

MMLU-Pro

IFEval

PhiBench

GPQA

LiveBench

SimpleQA

Model Comparison

See how Phi-4 stacks up against other leading models across key performance metrics.

84.8

MMLU - Phi-4

84.8

(94%)

85.9

MMLU - Gemini 1.5 Pro

85.9

(95%)

MMLU - Llama 3.3 70B Instruct

(96%)

87.5

MMLU - Grok-2

87.5

(97%)

83.3

MMLU - Qwen2.5 32B Instruct

83.3

(93%)

86.2

MMLU - Grok-2 mini

86.2

(96%)

56.1

GPQA - Phi-4

56.1

(62%)

59.1

GPQA - Gemini 1.5 Pro

59.1

(66%)

50.5

GPQA - Llama 3.3 70B Instruct

50.5

(56%)

56.0

GPQA - Grok-2

56.0

(62%)

49.5

GPQA - Qwen2.5 32B Instruct

49.5

(55%)

GPQA - Grok-2 mini

(57%)

80.4

MATH - Phi-4

80.4

(89%)

86.5

MATH - Gemini 1.5 Pro

86.5

(96%)

MATH - Llama 3.3 70B Instruct

(86%)

76.1

MATH - Grok-2

76.1

(85%)

83.1

MATH - Qwen2.5 32B Instruct

83.1

(92%)

MATH - Grok-2 mini

(81%)

82.6

HumanEval - Phi-4

82.6

(92%)

84.1

HumanEval - Gemini 1.5 Pro

84.1

(93%)

88.4

HumanEval - Llama 3.3 70B Instruct

88.4

(98%)

88.4

HumanEval - Grok-2

88.4

(98%)

88.4

HumanEval - Qwen2.5 32B Instruct

88.4

(98%)

85.7

HumanEval - Grok-2 mini

85.7

(95%)

70.4

MMLU-Pro - Phi-4

70.4

(78%)

75.8

MMLU-Pro - Gemini 1.5 Pro

75.8

(84%)

68.9

MMLU-Pro - Llama 3.3 70B Instruct

68.9

(77%)

75.5

MMLU-Pro - Grok-2

75.5

(84%)

MMLU-Pro - Qwen2.5 32B Instruct

(77%)

MMLU-Pro - Grok-2 mini

(80%)

MMLU

GPQA

MATH

HumanEval

MMLU-Pro

Phi-4

Gemini 1.5 Pro

Llama 3.3 70B Instruct

Grok-2

Qwen2.5 32B Instruct

Grok-2 mini

Detailed Benchmarks

Dive deeper into Phi-4's performance across specific task categories. Expand each section to see detailed metrics and comparisons.

Coding

HumanEval

93.7%

84.8%

84.8%

84.1%

83.5%

82.6%

81.1%

81.1%

Llama 3.1 70B Instruct

80.5%

Ministral 8B Instruct

34.8%

Current model

Other models

Avg (79.1%)

HumanEval+

Phi-4

82.8%

Qwen2.5 32B Instruct

52.4%

Qwen2.5 14B Instruct

51.2%

Current model

Other models

Avg (62.1%)

Reasoning

DROP

DeepSeek-R1

92.2%

Llama 3.1 70B Instruct

79.6%

79.3%

78.9%

78.4%

75.5%

74.9%

74.3%

70.2%

Llama 3.1 8B Instruct

59.5%

Current model

Other models

Avg (76.3%)

Knowledge

MMLU

91.8%

Llama 3.3 70B Instruct

86.0%

85.9%

85.9%

85.2%

84.8%

84.0%

84.0%

Llama 3.1 70B Instruct

83.6%

Llama 3.2 3B Instruct

63.4%

Current model

Other models

Avg (83.5%)

GPQA

87.7%

60.0%

59.4%

59.1%

59.1%

56.1%

56.0%

53.6%

53.6%

25.3%

Current model

Other models

Avg (57.0%)

MATH

97.9%

86.5%

85.5%

83.1%

83.1%

80.4%

80.0%

78.3%

78.0%

32.6%

Current model

Other models

Avg (78.5%)

Non categorized

MGSM

o3-mini

92.0%

Llama 3.2 90B Instruct

86.9%

85.6%

83.5%

82.6%

80.6%

75.1%

74.5%

Llama 3.2 11B Instruct

68.9%

Phi-3.5-mini-instruct

47.9%

Current model

Other models

Avg (77.8%)

SimpleQA

Grok-3

43.6%

42.6%

42.4%

30.1%

24.9%

13.8%

Mistral Small 3.1 24B

10.4%

Phi-4

3.0%

Current model

Other models

Avg (26.4%)

MMLU-Pro

DeepSeek-R1

84.0%

Llama 3.1 405B Instruct

73.3%

72.6%

72.0%

71.1%

70.4%

69.0%

Llama 3.3 70B Instruct

68.9%

Claude 3 Opus

68.5%

Qwen2.5-Coder 7B Instruct

40.1%

Current model

Other models

Avg (69.0%)

ArenaHard

DeepSeek-R1

92.3%

DeepSeek-V2.5

76.2%

Phi-4

75.4%

Current model

Other models

Avg (81.3%)

LiveBench

QwQ 32B

73.1%

67.0%

52.3%

52.3%

47.6%

35.9%

Current model

Other models

Avg (54.7%)

IFEval

QwQ 32B

83.9%

DeepSeek-R1

83.3%

Mistral Small 3

82.9%

Llama 3.1 8B Instruct

80.4%

Llama 3.2 3B Instruct

77.4%

Qwen2.5 7B Instruct

71.2%

Phi-4

63.0%

Pixtral-12B

61.3%

Current model

Other models

Avg (75.4%)

Providers Pricing Coming Soon

We're working on gathering comprehensive pricing data from all major providers for Phi-4. Compare costs across platforms to find the best pricing for your use case.

OpenAI

Anthropic

Google

Mistral AI

Cohere

Share your feedback

Hi, I'm Charlie Palars, the founder of Deepranking.ai. I'm always looking for ways to improve the site and make it more useful for you. You can write me through this form or directly through X at @palarsio.

Your feedback helps us improve our service

Phi-4

Model Specifications

Core Specifications

Capabilities & License

Resources

Performance Insights

Model Comparison

Detailed Benchmarks

Coding

HumanEval

HumanEval+

Reasoning

DROP

Knowledge

MMLU

GPQA

MATH

Non categorized

MGSM

SimpleQA

MMLU-Pro

ArenaHard

LiveBench

IFEval

Providers Pricing Coming Soon

Share your feedback

Stay Ahead with AI Updates