Qwen·Gemini Soccer Video Benchmark

On April 7, 2026, without any specialized training or domain tuning, the same prompt was run on a 118-second broadcast-quality single-camera clip to compare Gemini Pro, Gemini Free, and Qwen 3.6 Plus.

The ground truth data consisted of 16 shots.

• On-target 10
• Off-target 6
• Right foot 10
• Left foot 4
• Header 2

None of the models' results fully matched the ground truth.

• Gemini Pro: 10 shots, on-target 5, off-target 5, right foot 8, left foot 2, header 0
• Gemini Free: 12 shots, on-target 7, off-target 5, right foot 9, left foot 1, header 2
• Qwen 3.6 Plus: 9 shots, on-target 6, off-target 3, right foot 6, left foot 1, header 2

Gemini Free's timestamps were about 1 second later than the actual moments and it missed some shots. Qwen 3.6 Plus counted fewer shots, and Gemini Pro also had errors remaining.

The conclusion is that the gap between looking plausible and the accuracy needed for professional analysis is still large. Among the three models, Gemini Pro was the most stable, but clear errors still remained. However, since this was a single-clip experiment, the scope for generalization is limited.