Evaluation overview

Overall rank

#10

91 points · 201 models

Open-weight rank

#2

94.8 points · 119 models

Data coverage

25 items

4 sources · updated through 2026-09-02

Domain performance

Coding#9 / 181

91.6

Research#10 / 195

96.1

Agents#18 / 147

87.4

Coding · GLM-5.3 Official Release

84.5%

1/ 27

Reasoning · Arena.ai Text · Max mode

1,482.13Elo

18/ 375

Agents · GLM-5.3 Official Release

62.5%

3/ 47

Cybersecurity · GLM-5.3 Official Release

105tasks

3/ 7

Coding · Arena.ai WebDev · Max mode

1,609.17Elo

10/ 106

Agents · Arena.ai Agent · Max mode

12.36%

6/ 52

Context (input / output): 1.05M / 131K
Parameters: 744B / 40B active
Release date: 2026-08
License: GLM-5.3 License
Input modalities: Text
Output modalities: Text
Base model: Same as GLM-5.2 (all gains come from post-training)
Precision: FP8 (default) / BF16 (separate repository)
Architecture: MoE (256 experts), DSA, IndexShare
Exact release date: 2026-08-14
API model ID: glm-5.3
Availability: GLM Coding Plan, ZCode, and API
Reasoning mode: Thinking only (supports low / high / max)
Weight status: Open weights released (GLM-5.3 License)
Evaluation source: Official 16-row comparison table, split into 17 metrics and 120 scores
Default reasoning effort: max (supports low / high / max)