CogAgent
Zhipu AI (Z.ai) · 2023-12-15
CogAgent is an 18B visual-language model developed by Zhipu AI and Tsinghua University for GUI understanding and autonomous navigation. It extends CogVLM with a high-resolution cross-attention module for 1120×1120 inputs and can generate plans, actions, and screen coordinates. Its model weights are governed by the CogVLM License; Apache-2.0 applies to repository code rather than the weights.
Benchmark scores
| Benchmark | Score |
|---|---|
| Capability | 96.9 |
| CharXiv Reasoning | 18.8 |
| CharXiv-D | 36.3 |
| OSWorld | 1.1 |
| OSWorld (self-reported, Screenshot + A11y tree) | 1.3 |
| OSWorld (self-reported, Set-of-Mark) | 1.0 |
| ScreenSpot-Pro | 7.7 |