DeepSeek-V2.5

DeepSeek

DeepSeek-V2.5 represents a significant upgrade, merging the strengths of DeepSeek-V2-Chat and DeepSeek-Coder-V2-Instruct to deliver enhanced general and coding capabilities. This new version is more attuned to human expectations and incorporates optimizations across a range of areas, such as improved writing proficiency and more accurate instruction following.

Model Specifications

Technical details and capabilities of DeepSeek-V2.5

Core Specifications

236.0B Parameters

Model size and complexity

8.2K / 8.2K

Input / Output tokens

May 7, 2024

Release date

Capabilities & License

Multimodal Support

Not Supported

Web Hydrated

License

deepseek

Resources

Research Paper

https://arxiv.org/abs/2405.04434

API Reference

https://www.deepseek.com/

Playground

https://huggingface.co/deepseek-ai/DeepSeek-V2.5

Code Repository

https://huggingface.co/deepseek-ai/DeepSeek-V2.5

Performance Insights

Check out how DeepSeek-V2.5 handles various AI tasks through comprehensive benchmark results.

100

95.1

GSM8K

95.1

(95%)

90.2

MT-Bench

90.2

(90%)

HumanEval

(89%)

84.3

BBH

84.3

(84%)

80.4

AlignBench

80.4

(80%)

80.4

MMLU

80.4

(80%)

78.3

DS-FIM-Eval

78.3

(78%)

76.2

ArenaHard

76.2

(76%)

74.7

MATH

74.7

(75%)

73.8

HumanEval Multi

73.8

(74%)

72.2

Aider

72.2

(72%)

63.1

DS-Arena-Code

63.1

(63%)

50.5

AlpacaEval 2.0

50.5

(51%)

41.8

LiveCodeBench

41.8

(42%)

16.8

SWE-verified

16.8

(17%)

GSM8K

MT-Bench

HumanEval

BBH

AlignBench

MMLU

DS-FIM-Eval

ArenaHard

MATH

HumanEval Multi

Aider

DS-Arena-Code

AlpacaEval 2.0

LiveCodeBench

SWE-verified

Model Comparison

See how DeepSeek-V2.5 stacks up against other leading models across key performance metrics.

100

HumanEval - DeepSeek-V2.5

(89%)

HumanEval - Nova Pro

(89%)

88.4

HumanEval - Qwen2.5 32B Instruct

88.4

(88%)

83.5

HumanEval - Qwen2.5 14B Instruct

83.5

(84%)

81.1

HumanEval - Nova Micro

81.1

(81%)

87.8

HumanEval - Gemma 3 27B

87.8

(88%)

80.4

MMLU - DeepSeek-V2.5

80.4

(80%)

85.9

MMLU - Nova Pro

85.9

(86%)

83.3

MMLU - Qwen2.5 32B Instruct

83.3

(83%)

79.7

MMLU - Qwen2.5 14B Instruct

79.7

(80%)

77.6

MMLU - Nova Micro

77.6

(78%)

76.9

MMLU - Gemma 3 27B

76.9

(77%)

95.1

GSM8K - DeepSeek-V2.5

95.1

(95%)

94.8

GSM8K - Nova Pro

94.8

(95%)

95.9

GSM8K - Qwen2.5 32B Instruct

95.9

(96%)

94.8

GSM8K - Qwen2.5 14B Instruct

94.8

(95%)

92.3

GSM8K - Nova Micro

92.3

(92%)

95.9

GSM8K - Gemma 3 27B

95.9

(96%)

74.7

MATH - DeepSeek-V2.5

74.7

(75%)

76.6

MATH - Nova Pro

76.6

(77%)

83.1

MATH - Qwen2.5 32B Instruct

83.1

(83%)

MATH - Qwen2.5 14B Instruct

(80%)

69.3

MATH - Nova Micro

69.3

(69%)

MATH - Gemma 3 27B

(89%)

84.3

BBH - DeepSeek-V2.5

84.3

(84%)

86.9

BBH - Nova Pro

86.9

(87%)

84.5

BBH - Qwen2.5 32B Instruct

84.5

(85%)

78.2

BBH - Qwen2.5 14B Instruct

78.2

(78%)

79.5

BBH - Nova Micro

79.5

(80%)

87.6

BBH - Gemma 3 27B

87.6

(88%)

HumanEval

MMLU

GSM8K

MATH

BBH

DeepSeek-V2.5

Nova Pro

Qwen2.5 32B Instruct

Qwen2.5 14B Instruct

Nova Micro

Gemma 3 27B

Detailed Benchmarks

Dive deeper into DeepSeek-V2.5's performance across specific task categories. Expand each section to see detailed metrics and comparisons.

Math

GSM8K

97.1%

96.4%

95.9%

95.9%

95.8%

95.1%

95.0%

94.8%

94.8%

68.6%

Current model

Other models

Avg (92.9%)

Coding

HumanEval

93.7%

92.0%

92.0%

90.2%

89.0%

89.0%

Llama 3.1 405B Instruct

89.0%

Mistral Small 3.1 24B

88.4%

Llama 3.3 70B Instruct

88.4%

Ministral 8B Instruct

34.8%

Current model

Other models

Avg (84.7%)

LiveCodeBench

80.0%

63.4%

62.5%

55.5%

50.0%

41.8%

40.5%

39.0%

37.6%

Qwen2.5-Coder 7B Instruct

18.2%

Current model

Other models

Avg (48.9%)

Knowledge

MMLU

91.8%

Grok-1.5

81.3%

Jamba 1.5 Large

81.2%

Mistral Small 3.1 24B

80.6%

Nova Lite

80.5%

DeepSeek-V2.5

80.4%

Llama 3.1 Nemotron 70B Instruct

80.2%

Qwen2.5 14B Instruct

79.7%

Claude 3 Sonnet

79.0%

Llama 3.2 3B Instruct

63.4%

Current model

Other models

Avg (79.8%)

MATH

97.9%

76.6%

76.6%

76.1%

75.5%

74.7%

Llama 3.1 405B Instruct

73.8%

Nova Lite

73.3%

Grok-2 mini

73.0%

Gemini 1.0 Pro

32.6%

Current model

Other models

Avg (73.0%)

Non categorized

ArenaHard

DeepSeek-R1

92.3%

DeepSeek-V2.5

76.2%

Phi-4

75.4%

Current model

Other models

Avg (81.3%)

AlignBench

81.6%

80.4%

73.3%

72.1%

Current model

Other models

Avg (76.9%)

MT-Bench

Llama 3.1 Nemotron 70B Instruct

9.0

90.2%

86.3%

84.1%

83.5%

Ministral 8B Instruct

83.0%

Pixtral-12B

76.8%

Current model

Other models

Avg (2.0)

Aider

DeepSeek-V2.5

72.2%

Qwen2.5-Coder 7B Instruct

55.6%

Current model

Other models

Avg (63.9%)

BBH

87.6%

86.9%

84.5%

84.3%

82.4%

82.4%

79.5%

78.2%

70.6%

Current model

Other models

Avg (81.8%)

Providers Pricing Coming Soon

We're working on gathering comprehensive pricing data from all major providers for DeepSeek-V2.5. 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

DeepSeek-V2.5

Model Specifications

Core Specifications

Capabilities & License

Resources

Performance Insights

Model Comparison

Detailed Benchmarks

Math

GSM8K

Coding

HumanEval

LiveCodeBench

Knowledge

MMLU

MATH

Non categorized

ArenaHard

AlignBench

MT-Bench

Aider

BBH

Providers Pricing Coming Soon

Share your feedback

Stay Ahead with AI Updates