Gemma 4 31B
Gemma 4 31B scores 38.8 out of 100 overall, which puts it in a tie for rank 19–29 of 36 models on the AI Vision Index. It is strongest at Not being fooled (78) and weakest at Counting (0). It beats the average of all 36 models in one of the six categories. It was tested on a smaller sample than most models, so its margin of error is wider and its single-test scores are rough.
Overall score38.895% interval 33.4–46.6
Rank19–29of 36 models (tied)
Cost per 1,000 answers$1.94
Seconds per answer2.0 s
Thinking effortdefault
Maker Google
WeightsOpen weights
AccessAPI
Tested
Reading36
Counting0
Space, 3D and motion31
Spotting changes and errors48
Diagrams40
Not being fooled78
The bar is the score; the light band is the 95% interval. The overall score is the average of the six.
Reading
Reading a clock20
Reading a meter40
Eye chart20
Reading a label60
Reading a chart50
Test strips40
Adding up coins20
Traffic light from a car40
Counting
Counting objects0
Counting stacked goods0
Counting on a conveyor0
Space, 3D and motion
Which is nearer20
Which way it faces6
Same or mirror image0
Will it topple70
Match the render0
Which bay can I park in?78
Follow the hidden object45
Spotting changes and errors
Spot the change80
Shelf against plan20
Rendering glitch76
Check the invoice15
Diagrams
Follow the cable40
Follow the pipe40
Missing dimension40
Not being fooled
Absent objects100
Saying "I can't tell"100
Optical illusions10
Real or AI-made100
Scores are corrected for guessing: 0 is no better than a guess, 100 is always right. All models on every test
Neighbours on the leaderboard
Badge for your site
HTML
<a href="https://aivisionindex.com/models/gemma-4-31b/"><img src="https://aivisionindex.com/badge/gemma-4-31b.svg" alt="Gemma 4 31B: ranked #19–29 of 36 on the AI Vision Index"></a>Markdown
[](https://aivisionindex.com/models/gemma-4-31b/)Results are CC BY 4.0, so you are free to reuse them if you credit the AI Vision Index. All results (JSON)