InSight-doc-8B vs Qwen3-VL-8B

Trajectories are replayed, not generated in real time. Timing is recreated 1:1 from actual generation under the same generation settings for both models, except that r=0.7 (high resolution) is used for Qwen3-VL-8B and r=0.35 (medium-low resolution) is used for InSight-doc-8B to match their overall accuracies on the benchmarks.

PDF Pages / Images (Downsampled)

Click any page thumbnail to view the original image.

Loading pages…

Question

Loading examples…

Qwen3-VL-8B

r=0.7 idle

InSight-doc-8B

r=0.35 idle