← All models

DeepSeek: DeepSeek V4 Pro

DeepSeek

open_source_flagship

Capabilities

ToolsStructured OutputExtended ThinkingLong ContextCode

Strengths

  • Exceptional 1M-token context window for long documents
  • Open weights enable full data privacy and self-hosting
  • Strong multi-turn conversation and reasoning capabilities
  • Efficient MoE architecture balancing quality and speed

Weaknesses

  • Extremely slow inference speed (34 tok/s) limits real-time applications
  • No vision/multimodal capabilities
  • Limited transparency on training data composition
  • Reasoning features add latency overhead

Pricing

Input / 1M tokens

$0.435

Output / 1M tokens

$0.87

Context window

1,048.576k

Transparency

Open weights

10.0 / 10

Open training data

0.0 / 10

Open methodology

6.0 / 10

Licence openness

9.0 / 10

Provider disclosure

6.5 / 10

FMTI company score

N/A

Composite:6.5 / 10

Open-weight model enabling self-hosting and full data control. DeepSeek publishes model weights and some technical details, but training data sources not fully disclosed. MoE architecture and reasoning capabilities documented in research papers.

Sustainability

Inference energy

N/A

Training footprint

N/A

Provider infrastructure

5.0 / 10

Composite:5.0 / 10

DeepSeek operates in China with limited public disclosure of energy infrastructure. Open-weight design allows efficient self-hosted deployment. No official sustainability commitments published. MoE architecture with 49B activated parameters suggests moderate inference efficiency compared to dense alternatives.

Task Fitness

question answering
88%
code
82%
mathematics
84%
business communication
78%
analysis
85%
extraction
85%
long-form generation
92%
research
60%
reasoning
85%
summarisation
88%
translation
84%
conversation
93%