DeepSeek

DeepSeek-V3

Zero-eval
#1DROP
#1HumanEval-Mul
#1Aider-Polyglot Edit
+5 more

by DeepSeek

+
+
+
+
About

DeepSeek-V3 was introduced as a major architectural advancement, developed with 671B mixture-of-experts parameters and trained on 14.8 trillion tokens. Built to be three times faster than V2 while maintaining open-source availability, it demonstrates competitive performance against frontier closed-source models and represents a significant leap in efficient large-scale model design.

+
+
+
+
Pricing Range
Input (per 1M)$0.27 -$0.27
Output (per 1M)$1.10 -$1.10
Providers1
+
+
+
+
Timeline
AnnouncedDec 25, 2024
ReleasedDec 25, 2024
+
+
+
+
Specifications
Training Tokens14.8T
+
+
+
+
License & Family
License
MIT + Model License (Commercial use allowed)
Performance Overview
Performance metrics and category breakdown

Overall Performance

20 benchmarks
Average Score
67.2%
Best Score
91.6%
High Performers (80%+)
8

Performance Metrics

Max Context Window
262.1K
Avg Throughput
100.0 tok/s
Avg Latency
1ms
+
+
+
+
All Benchmark Results for DeepSeek-V3
Complete list of benchmark scores with detailed information
DROP
text
0.92
91.6%
Self-reported
CLUEWSC
text
0.91
90.9%
Self-reported
MATH-500
text
0.90
90.2%
Self-reported
MMLU-Redux
text
0.89
89.1%
Self-reported
MMLU
text
0.89
88.5%
Self-reported
C-Eval
text
0.86
86.5%
Self-reported
IFEval
text
0.86
86.1%
Self-reported
HumanEval-Mul
text
0.83
82.6%
Self-reported
Aider-Polyglot Edit
text
0.80
79.7%
Self-reported
MMLU-Pro
text
0.76
75.9%
Self-reported
Showing 1 to 10 of 20 benchmarks
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+