OpenGVLab: InternVL3 2B

OpenGVLab's InternVL3 2B, created on April 30, 2025, is positioned as a high-speed, reasonably performing multimodal large language model. It consistently ranks among the fastest models, demonstrating exceptional inference speed. Furthermore, it offers competitive pricing, ranking in the 89th percentile across benchmarks. While excelling in speed and cost-efficiency, the model exhibits significant variability in performance across different categories. In Classification, specifically Email Classification, it achieved 79.0% accuracy, indicating a reasonable ability to categorize structured text. However, its performance in more complex cognitive tasks is notably weaker. The model scored 0.0% accuracy in Reasoning, suggesting a current inability to handle multi-step logical problems. Similarly, in Ethics and General Knowledge, it achieved only 27.0% and 15.0% accuracy respectively, placing it in the lower percentiles for these domains. The longer durations observed in these benchmarks (Ethics: 225814ms, General Knowledge: 439530ms) further highlight the challenges it faces with these complex tasks. Its core strength lies in its speed and cost-effectiveness, making it potentially suitable for high-throughput, less cognitively demanding classification tasks. However, its current limitations in reasoning, ethical understanding, and general knowledge are significant weaknesses that would need to be addressed for broader application.