Comparison of Text-Only Benchmark Performance with Qwen2.5-Instruct

[y-rok]

1. Is this model fine-tuned on the Qwen2.5 model or the Qwen2.5-instruct model?

2. How much does the Qwen2.5-VL model's performance on text-based benchmarks degrade compared to the Qwen2.5-instruct model, which was trained solely on text? I am asking this question with the expectation that performance may decline due to training on vision-related data.

[DietDietDiet]

same question here. Anyone tested qwen2.5-vl on text benchmarks?