Kimi announced the open-source release of PerceptionBench, a multimodal model vision perception benchmark, on July 29. According to PANews, the benchmark breaks visual perception into 10 atomic abilities for separate evaluation, covering visual relationships, counting, attributes, depth and 3D, localization, comparison, fine-grained recognition, context integration, OCR, and hallucination detection.
The benchmark was built from model failure cases in 42 existing evaluation sets and includes 3,000 manually verified questions. Each question tests only one visual ability and does not require reasoning or external knowledge. Evaluation results showed that none of the 16 leading multimodal models achieved an overall accuracy above 60%.
AI TRENDS | Kimi Releases Open-Source Multimodal Vision Benchmark PerceptionBench
2026-07-28 16:03:40
Disclaimer:
1. The information provided does not constitute investment advice. Investors should make independent decisions and bear all risks themselves.
2. The copyright of this content belongs to the original author. The views expressed herein are solely those of the author and do not represent the stance or position of this website.
