OpenGVLab

/

InternVL3-2B-Instruct

like 3

Follow

OpenGVLab 998

Hi,

I've been testing the recently uploaded InvernVL3 models, specifically the InvernVL3-2B-Pretrain and InvernVL3-2B-Instruct versions.

I ran evaluations using VLMEvalKit on MMMU, VideoMME, and OCRBench, and got identical results across all three benchmarks for both models — the exact same scores in every case.

This makes me wonder if there might have been an issue with the uploads or checkpoint configurations, and whether both versions might actually point to the same underlying model.

Is this expected behavior?

Thanks in advance — any clarification would be greatly appreciated.