Name: DeepSeek-V4-Pro-0813
Author: DeepSeek

Evaluation overview

Overall rank

#7

93.6 points · 194 models

Open-weight rank

#2

96.1 points · 116 models

Data coverage

13 items

2 sources · updated through 2026-08-14

Domain performance

Coding#8 / 174

91.6

Research#14 / 188

94.5

Agents#9 / 140

88.7

Coding · DeepSeek-V4-Pro-0813 Official Release

61.5%

2/ 34

Agents · GLM-5.3 Official Release

60%

4/ 45

Reasoning · DeepSeek-V4-Pro-0813 Official Release

42.7%

11/ 67

Coding · GLM-5.3 Official Release

83.3%

4/ 26

Agents · GLM-5.3 Official Release

87.9%

5/ 26

Agents · GLM-5.3 Official Release

43.2%

5/ 22

Context (input / output): 1.05M / 393K
Parameters: 1650.5B / 49B active
Release date: 2026-08
License: MIT
Input modalities: Text
Output modalities: Text
Architecture: MoE (384 routed experts), DSA + DSpark
Modes: Supports non-thinking and thinking (low / high / max)
Version: 0813
Precision: FP4 MoE experts + FP8
Total parameters: Approx. 1.65T
Exact release date: 2026-08-13
Experts: 384 (6 routed + 1 shared expert per token)
Release status: Production release (app / web / API)
Weight status: Official 0813 weights released (MIT)
Evaluation source: Official complete comparison table with 11 split agent metrics
API model ID: deepseek-v4-pro
Responses API: Native support, with dedicated Codex compatibility