Atlas

Models

← All models

CogAgent-VQA

Zhipu AI (Z.ai) · 2023-12-15

CogAgent-VQA is Zhipu AI and Tsinghua University's original CogAgent checkpoint optimized for single-turn visual question answering. It shares the 18B-class CogAgent architecture with a high-resolution visual cross-attention module, but has different released weights from the chat variant used for GUI navigation and multi-turn dialogue. Its weights are governed by the CogVLM License; CharXiv's official evaluation implementation explicitly selects this VQA checkpoint.

Benchmark scores

BenchmarkScore
CharXiv Reasoning18.8
CharXiv-D36.3
Loading Atlas data…