SWE-MMMultimodalCodingAgents

SWE-MM corresponds to SWE-bench Multimodal. It uses real software issues containing screenshots, design mockups, diagrams, or visual error information to evaluate whether models can combine textual and visual context to generate a resolving patch.

Ranking Sources

#ModelScore

1

Qwen3.8-27BAlibaba

38.6%

2

Qwen3.7-PlusAlibaba

30.0%

3

Claude Opus 4.6Anthropic

27.1%

4

Qwen3.6-27BAlibaba

25.7%

4 models total

Back to benchmarks