DeepSeek V3.1

DeepSeek

balanced

Capabilities

ToolsCodeStructured OutputMultilingual

Strengths

Extraordinary cost/performance
Strong coding
Open weights
MIT licensed

Weaknesses

No vision
128K context
China-based data jurisdiction

Pricing

Input / 1M tokens

$0.21

Output / 1M tokens

$0.7899999999999999

Context window

128k

Transparency

Open weights

10.0 / 10

Open training data

3.0 / 10

Open methodology

7.0 / 10

Licence openness

10.0 / 10

Provider disclosure

2.0 / 10

FMTI company score

300.0 / 10

Composite:6.4 / 10

MIT licence + open weights. DeepSeek FMTI 30.

Sustainability

Inference energy

6.5 / 10

Training footprint

N/A

Provider infrastructure

2.0 / 10

Composite:4.3 / 10

MoE architecture with 93.3% memory reduction. Only 37B active params of 671B total. China grid heavily coal-based, no provider sustainability reporting.

Task Fitness

question answering

79%

code

90%

mathematics

86%

business communication

66%

analysis

82%

extraction

85%

long-form generation

78%

research

57%

reasoning

82%

summarisation

82%

translation

82%

conversation

75%