RSI-Jev v4.0-VL · Gallery

What a 2B model decides from a picture

RSI-Jev v4.0-VL takes an image and a question with fixed answers, and returns a probability for each answer in one pass, with no text generation. Every result below was run on v4.0-VL on one GB10 desktop GPU. Right and wrong answers are both shown.

86.6

VisA defect detection, AUROC, never shown a defect

7/24

Breakout bricks cleared, 3 lives left, stopped at 200 decisions

45/55

answers right on the cards shown (187/227 across all 143 items we ran, on 89 images)

62 ms

median time per answer on the GB10 over all 227, including image preprocessing

Factory inspection

Photos of real parts from the VisA dataset. The model has never seen a defect example: it gets one photo and says whether the part is good or defective.

All 2,162 VisA test photos, 12 part types, scored by area under the ROC curve (AUROC): 50 is a coin flip, 100 is perfect. 86.6 for RSI-Jev v4.0-VL, +5.5 points vs Jev-Omni and +3.7 vs Gemma 4 12B on the same photos. The score ranks parts well, but it is cautious: taken at a plain 50% line it flags 405 of 1,200 defective parts and 34 of 962 good ones, so a real line would set its own threshold.

System	AUROC	95% range	Saw good examples first?
RSI-Jev v4.0-VL (2B)	86.6	85.1–88.1	no
PatchCore, 16 good parts per category	85.7	84.2–87.2	16 good parts
Gemma 4 12B (base)	82.9	81.3–84.5	no
Jev-Omni (Gemma 4 12B + decision head)	81.1	79.4–82.7	no

capsules: defect

Production-line inspection photo of green gel capsules.

Is everything in the photo good, or is at least one part defective?

Defective68%

All good32%

✓ right254 ms

Reference answer: Defective

Same photo, defect score with both answer orders averaged: this model 75%, Jev-Omni (12B) 100%, Gemma 4 12B 100%.

Source and license

VisA (Zou et al., ECCV 2022), test split. Amazon Science, VisA dataset. License: CC BY 4.0.

capsules: good part

Production-line inspection photo of green gel capsules.

Is everything in the photo good, or is at least one part defective?

All good91%

Defective9%

✓ right250 ms

Reference answer: All good

Same photo, defect score with both answer orders averaged: this model 9%, Jev-Omni (12B) 76%, Gemma 4 12B 72%.

Source and license

VisA (Zou et al., ECCV 2022), test split. Amazon Science, VisA dataset. License: CC BY 4.0.

pcb4: defect

Production-line inspection photo of a printed circuit board.

Is everything in the photo good, or is at least one part defective?

Defective76%

All good24%

✓ right245 ms

Reference answer: Defective

Same photo, defect score with both answer orders averaged: this model 81%, Jev-Omni (12B) 82%, Gemma 4 12B 90%.

Source and license

VisA (Zou et al., ECCV 2022), test split. Amazon Science, VisA dataset. License: CC BY 4.0.

pcb4: good part

Production-line inspection photo of a printed circuit board.

Is everything in the photo good, or is at least one part defective?

All good86%

Defective14%

✓ right239 ms

Reference answer: All good

Same photo, defect score with both answer orders averaged: this model 17%, Jev-Omni (12B) 68%, Gemma 4 12B 20%.

Source and license

VisA (Zou et al., ECCV 2022), test split. Amazon Science, VisA dataset. License: CC BY 4.0.

candle: defect

Honest miss

Production-line inspection photo of candles.

Is everything in the photo good, or is at least one part defective?

All good94%

Defective6%

✗ wrong239 ms

Reference answer: Defective

Missed: a chunk of wax is missing from the top-left candle. Same photo, defect score with both answer orders averaged: this model 7%, Jev-Omni (12B) 65%, Gemma 4 12B 36%.

Source and license

VisA (Zou et al., ECCV 2022), test split. Amazon Science, VisA dataset. License: CC BY 4.0.

chewinggum: good part

Honest miss

Production-line inspection photo of a piece of chewing gum.

Is everything in the photo good, or is at least one part defective?

Defective87%

All good13%

✗ wrong245 ms

Reference answer: All good

False alarm on a good piece. Same photo, defect score with both answer orders averaged: this model 86%, Jev-Omni (12B) 42%, Gemma 4 12B 7%.

Source and license

VisA (Zou et al., ECCV 2022), test split. Amazon Science, VisA dataset. License: CC BY 4.0.

Road signs and lights

Photos from the vision-jev driving game. Its own questions: what should the car do, and what is this thing?

Shown here: 11 of 13 answers right. Whole pool we ran for this section: 99 of 124 (60 pictures).

stop, in our lane

What should the car do about the object in the image, which is on the road, in the car's lane, ahead of the car?

stop then go91%

stop4%

wait for green4%

slow down<1%

continue<1%

2 others<1%

✓ right76 ms

Reference answer: stop then go

What kind of thing is the object in the image? stop sign 99% ✓ right

Source and license

Wikimedia Commons, "Stop sign us.jpg" (via reinhard-z/vision-jev samples, MIT). Dori. License: Public domain.

red, in our lane

What should the car do about the object in the image, which is on the road, in the car's lane, ahead of the car?

wait for green97%

stop then go2%

stop<1%

go<1%

continue<1%

2 others<1%

✓ right65 ms

Reference answer: wait for green

What kind of thing is the object in the image? traffic light 99.9% ✓ right

Source and license

Wikimedia Commons, "Red traffic signal, Stamford Road, Singapore - 20111210-01.jpg" (via reinhard-z/vision-jev samples, MIT). Jacklee. License: CC BY-SA 3.0.

green, in our lane

What should the car do about the object in the image, which is on the road, in the car's lane, ahead of the car?

go96%

continue2%

stop then go<1%

wait for green<1%

slow down<1%

2 others<1%

✓ right65 ms

Reference answer: go or continue

What kind of thing is the object in the image? traffic light 99.9% ✓ right

Source and license

Wikimedia Commons, "Green traffic signal, Stamford Road, Singapore - 20111210-03.jpg" (via reinhard-z/vision-jev samples, MIT). Jacklee. License: CC BY-SA 3.0.

speed limit 80, in our lane

What should the car do about the object in the image, which is on the road, in the car's lane, ahead of the car?

change speed98%

stop then go<1%

continue<1%

go<1%

slow down<1%

2 others<1%

✓ right65 ms

Reference answer: change speed

What kind of thing is the object in the image? speed limit sign 99.9% ✓ right
Which speed limit number is written on the object in the image? 80 99.9% ✓ right

Source and license

Wikimedia Commons, "Korean Sign - Maximum Speed Limit 80kph 1.jpg" (via reinhard-z/vision-jev samples, MIT). P.Ctnt. License: Public domain.

child, in our lane

Unsure

What should the car do about the object in the image, which is on the road, in the car's lane, ahead of the car?

stop28%

continue19%

go17%

slow down17%

wait for green12%

2 others8%

✓ right69 ms

Reference answer: stop

What kind of thing is the object in the image? person 37% ✓ right

Right, but only 28% sure. For people, animals and objects in the scene, the model often spreads its answer across stop, slow down and continue.

Source and license

Wikimedia Commons, "Boy and Ball (8257295231).jpg" (via reinhard-z/vision-jev samples, MIT). Evgeniy Isaev. License: CC BY 2.0.

leaves, in our lane

Honest miss

What should the car do about the object in the image, which is on the road, in the car's lane, ahead of the car?

stop43%

stop then go17%

continue16%

slow down10%

go8%

2 others5%

✗ wrong64 ms

Reference answer: continue

What kind of thing is the object in the image? other sign 73% ✗ wrong

Wrong: a pile of leaves can be driven over. The model says stop.

Source and license

Wikimedia Commons, "Laubdeponie, Gartenanlage Steinhof, Wien.JPG" (via reinhard-z/vision-jev samples, MIT). Wald1siedel. License: CC BY-SA 3.0.

Breakout from pixels

The jev-visual Breakout game, played live. Each frame goes to the model, which picks the lane the ball is in.

RSI-Jev v4.0-VL playing Breakout: the board and the decision feed with lane probabilities

Bricks cleared: 7 of 24
Lives left: 3 of 3
Ball returns: 6
Decisions: 200
Play time: 33 s
How it ended: stopped by the game's own 200-decision limit (not a win)
Model time per frame: 75 ms (median)
Ten-scene lane check: 10 of 10

The game is hr98w/jev-visual (MIT), unchanged. It sends one 640×480 screenshot and asks which of five columns holds the ball; the paddle then moves toward that column. The model gets no coordinates. The game stops the AI after 200 decisions. At about 75 ms per frame that came after 33 seconds of play, with no ball lost. The game was not finished.

Driving game, live

The Jev Driver game (reinhard-z/vision-jev, MIT), unmodified, with RSI-Jev v4.0-VL deciding from each dropped photo instead of from a caption. One live run, nine decisions.

6 of 9 decisions match the game's own expected actions. 121–174 ms per decision (median 130) on one GB10.

Photos in the clip, from Wikimedia Commons via the game: 80 sign by P.Ctnt (public domain); dog by Joselodos, teddy by Pixabay and adult by Peter Miranda (CC0); child by Evgeniy Isaev (CC BY 2.0); red and green lights by Jacklee and leaves by Wald1siedel (CC BY-SA 3.0). The 30 sign is the game's own sample, whose source it does not record. The sidebar also shows thumbnails of the game's other samples, among them amber light (Jacklee, CC BY-SA 3.0), cat (Grendelkhan) and box (MrBeastRapper, both CC BY-SA 4.0). Frames showing CC BY-SA photos are shared under the same licence. Full credits.

Right: 30 and 80 signs: change speed · red light: wait for green · green light: go · dog beside the lane: slow down · teddy in the lane: stop
Wrong: child in the car's lane: go (25%), so the car drives through · adult on the far sidewalk: slow down (expected continue)
Over-cautious: pile of leaves in the lane: stop (26%); the car waits until it is cleared

The worst miss of the run is in the clip: with a child in its lane the model chose go, at low confidence. Do not use it to drive anything.

How this run differs from the game's own pipeline: the Florence-2 caption step is bypassed, and the game's downscaled image, its location sentence and its three questions go to RSI-Jev, which answers in place of Jev. The game's labels still say "Jev". Recorded live on one GB10.

Is it really there?

Yes/no questions about objects that are, and are not, in the picture. Several absent objects usually appear with the ones that are there.

Shown here: 15 of 15 answers right. Whole pool we ran for this section: 26 of 26 (9 pictures).

dog photo: what is really there?

Is there a dog in the image?

Yes99.9%

No<1%

✓ right56 ms

Reference answer: Yes

Is there a cat in the image? No 99.9% ✓ right
Is there a frisbee in the image? No 99.9% ✓ right

Source and license

Wikimedia Commons, "Boxer dog posing on the grass.jpg" (via reinhard-z/vision-jev samples, MIT). Joselodos. License: CC0.

car photo: what is really there?

Is there a car in the image?

Yes99.9%

No<1%

✓ right52 ms

Reference answer: Yes

Is there a wind turbine in the image? Yes 99.9% ✓ right
Is there a person in the image? No 99.9% ✓ right

Source and license

Wikimedia Commons, "Public domain image - Peugeot iOn electric car in front of wind turbines.JPG" (via reinhard-z/vision-jev samples, MIT). Kiwiev. License: CC0.

teddy photo: what is really there?

Is there a teddy bear in the image?

Yes99.9%

No<1%

✓ right60 ms

Reference answer: Yes

Is there a bench in the image? Yes 99.9% ✓ right
Is there a child in the image? No 99.9% ✓ right

Source and license

Wikimedia Commons, "Heliograph of Bear ... .png" (via reinhard-z/vision-jev samples, MIT). Pixabay. License: CC0.

adult photo: what is really there?

Is there a person in the image?

Yes99%

No1%

✓ right51 ms

Reference answer: Yes

Is there a car in the image? Yes 80% ✓ right
Is there an umbrella in the image? No 94% ✓ right

Our first reference said "no car". The model said yes, and it is right: cars are visible down the street.

Source and license

Wikimedia Commons, "Pedestrians using crosswalk (Unsplash).jpg" (via reinhard-z/vision-jev samples, MIT). Peter Miranda. License: CC0.

Shapes: what is really there?

Is there a triangle in the image?

No99.9%

Yes<1%

✓ right116 ms

Reference answer: No

Is there a blue square in the image? Yes 98% ✓ right
Is there anything green in the image? No 99.9% ✓ right

Source and license

Synthetic, generated for this gallery. RSI-Jev authors. License: CC0 (generated).

Counting and positions

Rendered scenes from CLEVR and our own shape pictures: count things, compare them, find where they are.

Shown here: 4 of 6 answers right. Whole pool we ran for this section: 37 of 44 (44 pictures).

Count the red circles

How many red circles are there?

465%

330%

53%

6<1%

2<1%

3 others<1%

✓ right118 ms

Reference answer: 4

Source and license

Synthetic, generated for this gallery. RSI-Jev authors. License: CC0 (generated).

Count the orange squares

Unsure

How many orange squares are there?

752%

843%

64%

5<1%

4<1%

3 others<1%

✓ right118 ms

Reference answer: 7

Source and license

Synthetic, generated for this gallery. RSI-Jev authors. License: CC0 (generated).

Count the green triangles

Honest miss

How many green triangles are there?

270%

327%

41%

1<1%

5<1%

3 others<1%

✗ wrong117 ms

Reference answer: 3

Wrong: there are three green triangles; it says two.

Source and license

Synthetic, generated for this gallery. RSI-Jev authors. License: CC0 (generated).

CLEVR scene

Are there any blue objects to the left of the red cylinder?

No85%

Yes15%

✓ right50 ms

Reference answer: No

Source and license

CLEVR v1.0 validation (via HuggingFaceM4/the_cauldron, clevr subset). Johnson et al., CLEVR (2017). License: CC BY 4.0.

CLEVR scene

Is the number of gray cubes greater than the number of large purple rubber cylinders?

Yes81%

No19%

✓ right48 ms

Reference answer: Yes

Source and license

CLEVR v1.0 validation (via HuggingFaceM4/the_cauldron, clevr subset). Johnson et al., CLEVR (2017). License: CC BY 4.0.

CLEVR scene

Honest miss

Are there more large purple balls than balls?

Yes90%

No10%

✗ wrong49 ms

Reference answer: No

Source and license

CLEVR v1.0 validation (via HuggingFaceM4/the_cauldron, clevr subset). Johnson et al., CLEVR (2017). License: CC BY 4.0.

Charts, diagrams and documents

Charts, a diagram, game boards and receipts that we generated. The model reads each picture and makes a decision.

Shown here: 5 of 7 answers right. Whole pool we ran for this section: 13 of 15 (14 pictures).

Pie chart: biggest share

Which channel brings the most traffic?

Direct99.9%

Social<1%

Search<1%

Ads0%

Email0%

✓ right132 ms

Reference answer: Direct

Source and license

Synthetic, generated for this gallery. RSI-Jev authors. License: CC0 (generated).

Two lines: which ends cheaper?

Which plan is cheaper in the last month shown?

B94%

A6%

✓ right117 ms

Reference answer: B

Source and license

Synthetic, generated for this gallery. RSI-Jev authors. License: CC0 (generated).

Diagram: food chain

Which organism is the producer?

Grass99.9%

Rabbit0%

Eagle0%

Fox0%

✓ right80 ms

Reference answer: Grass

What does the fox eat? Rabbit 99.9% ✓ right

Source and license

Synthetic, generated for this gallery. RSI-Jev authors. License: CC0 (generated).

Is the axis honest?

The vertical axis starts at zero.

No98%

Yes2%

✓ right130 ms

Reference answer: No

Source and license

Synthetic, generated for this gallery. RSI-Jev authors. License: CC0 (generated).

Receipt vs policy: $38 + wine

Honest miss

Expense policy: meals up to $60 are approved automatically; alcohol is never reimbursable; anything else needs review.

What should happen to this expense?

approve81%

review12%

reject7%

✗ wrong88 ms

Reference answer: reject

Missed: the policy says alcohol is never paid back, and the receipt has a bottle of wine.

Source and license

Synthetic, generated for this gallery. RSI-Jev authors. License: CC0 (generated).

Tic-tac-toe: block

Honest miss

You play X. It is your move.

Which numbered cell must X take to stop O from winning?

239%

826%

618%

911%

36%

✗ wrong84 ms

Reference answer: 6

Wrong: O threatens the middle row, so X must take 6.

Source and license

Synthetic, generated for this gallery. RSI-Jev authors. License: CC0 (generated).

Screens and agent steps

Screens an agent sees: forms, dialogs, errors, and before/after screenshots of one step.

Shown here: 6 of 8 answers right. Whole pool we ran for this section: 8 of 12 (10 pictures).

Destructive dialog

The user said: "Actually, keep that file."

Which button should the agent click?

cancel99%

delete<1%

✓ right79 ms

Reference answer: cancel

The screen says the action can be undone. No 98% ✓ right

Source and license

Synthetic, generated for this gallery. RSI-Jev authors. License: CC0 (generated).

Payment declined

Goal: finish paying for the order.

What happened to the payment?

declined99.9%

pending<1%

paid<1%

✓ right121 ms

Reference answer: declined

What should the agent do next? another card 78% ✓ right

Source and license

Synthetic, generated for this gallery. RSI-Jev authors. License: CC0 (generated).

Alert dialog (high severity)

How urgently does an on-call engineer need to act on this?

high87%

medium10%

low3%

✓ right76 ms

Reference answer: high

Source and license

Synthetic, generated for this gallery. RSI-Jev authors. License: CC0 (generated).

Checkout form, 3 of 3 fields filled

Goal: finish paying for the order.

Which element should the agent use next to reach the goal?

pay99.9%

card number<1%

expiry<1%

cancel<1%

email0%

✓ right119 ms

Reference answer: pay

Source and license

Synthetic, generated for this gallery. RSI-Jev authors. License: CC0 (generated).

Checkout form, 1 of 3 fields filled

Honest miss

Goal: finish paying for the order.

Which element should the agent use next to reach the goal?

pay99%

card number1%

expiry<1%

cancel<1%

email0%

✗ wrong127 ms

Reference answer: card number

Confident and wrong: it jumps to Pay while the card number is still empty.

Source and license

Synthetic, generated for this gallery. RSI-Jev authors. License: CC0 (generated).

Before / after an agent step (layout broke)

Honest miss

Before deploy: After deploy:

The page layout changed or broke between the two screenshots.

No72%

Yes28%

✗ wrong216 ms

Reference answer: Yes

Missed: a large "Sign up!" box now covers the text, and the model says nothing broke.

Source and license

Synthetic, generated for this gallery. RSI-Jev authors. License: CC0 (generated).

Does 80% mean 80%?

How often the model is right when it is 60%, 70% or 90% sure.

v4.0-VL, MMStar + BLINK questions (3399 answers)

Each dot groups answers by how sure the model was. Dots on the dashed line mean its confidence matches how often it is right. Bigger dots hold more answers. Hover for numbers.

Test	Model	Qs	Right	Avg. sure	Gap (ECE)	Wrong at ≥99%
MMStar	v4.0-VL	1498	62.4%	73.2%	0.108	1
BLINK	v4.0-VL	1901	56.3%	63.4%	0.071	0

Gap (ECE) is the average distance from the dashed line; lower is better.

Model: RSI-Jev v4.0-VL (Qwen3.5-2B base with a trained decision head), bf16 with its built-in calibration, exactly as the server returns it. ◆ marks the reference answer. “Unsure” means the top answer got under 60%. Latency is the whole call for one answer on a shared GB10, warm, one at a time. Image sources and licenses are on each card; synthetic pictures were generated for this page. Built 2026-09-30 10:27.