Qwen vs Gemini: Same Screen Test, Which One Made It Up?

📺 Video (in Turkish) of this exact test: Çin'in Qwen'i Gemini'yi Geçti mi? Aynı Ekranı İkisine Verdim

Quick answer: In this Qwen vs Gemini test I gave Qwen 3.8 and Gemini 3.5 the same screenshot, first in full and then with the line holding the answer cropped out. On the full screen both were correct. Once cropped, Qwen invented "500 ml" on a bottle screen, and both models said "the customer received it" on a shipping screen. One check question fixed all three errors (our own test, 12 November 2026).

This article is the written summary of that test for anyone comparing Qwen vs Gemini for screenshot reading. The screens were mock-ups, the model answers were real. No third-party content was used and the video is not sponsored.

Two models took part:

• Qwen 3.8: Qwen3.8-27B with image input, called through Cloudflare Workers AI.
• Gemini 3.5: Gemini 3.5 Flash-Lite, called through Google AI for Developers.

Setup: 3 mock screens × full and cropped view × 2 rounds, plus 6 check questions.

How was the Qwen vs Gemini test kept fair?

Both models received the same screenshot and the same question. First the full screen was shown. Then the line that carried the answer was cropped and the same question was asked again.

The goal was simple: when the information is missing, does the model say so, or does it fill the gap with a guess? That difference matters for anyone putting AI into a business workflow. An agent that reports "sent" or "delivered" without real evidence leads to wrong decisions.

What did Qwen and Gemini get right on the full screen?

On the full screen both models answered correctly. When the information was clearly visible, both read it and reported it.

So reading ability was not the problem. The real test started when information was missing.

Which model made something up once the screen was cropped?

On the bottle screen, with the answer line cropped, Qwen said "500 ml". That number was not in the image; the model filled the gap itself. According to the video chapter titles, Gemini answered "information missing" in the same situation.

On the shipping screen both models made the same mistake: in the cropped view both said "the customer received it", even though the screen did not show who the parcel was delivered to.

Source: video description and our own test, 12 November 2026.

Can one check question fix the errors?

Yes, in this test it did. A single check question reversed all three errors.

The video also has a separate chapter testing whether asking openly for evidence helps. The exact wording of the check question and the models’ second answers are shown on screen in the video.

The practical lesson: never treat a model’s first answer as final. For numbers, dates, delivery status and payments, add a second verification step.

What does this mean for e-commerce automation?

If you let an AI agent read order panels, tracking pages or invoices, this test applies to you directly. A model saying "the customer received it" could start a return window early or trigger the wrong message to a buyer.

The issue is not intelligence. It is that the model guesses instead of staying silent when data is missing. So split your automation into two steps:

• Read step: the model extracts the information and quotes the line it came from.
• Decision step: if there is no evidence line, the flow stops and asks a person.

A flow built this way makes it much harder for a made-up answer to turn into real damage, whichever model you use. I covered a similar "don’t let AI invent facts" workflow in the AI product description generator guide.

Which one should you trust, and what should you check on your own screens?

This test does not prove that one model is better overall. It only covers these two model versions, these 3 mock screens and API answers from 12 November 2026. Results may differ on other screens or versions.

What it does show: both models guessed at least once when information was missing. What you should trust is your check step, not the model.

1. Send the full image. Cropping removes information and pushes the model to guess.
2. Ask for evidence. Add "quote the exact line your answer is based on" to your prompt.
3. Double-check status data. Do not leave delivery, payment or stock decisions to a single model answer.

For more side-by-side model tests, see Tencent’s free 295B model test and Fusion AI: 5 models vs GPT-5.5.

Frequently Asked Questions

Which was more accurate in the Qwen vs Gemini test?

Both were correct on the full screen. In the cropped view Qwen made up a number on the bottle screen, and both models made the same mistake on the shipping screen.

Why did Qwen say "500 ml"?

The line holding the answer was cropped, so the information was not in the image. Instead of saying it was missing, the model guessed a number.

How many screens were used?

3 mock screens, full and cropped views, 2 rounds and 6 check questions (12 November 2026).

Which model versions were tested?

Qwen3.8-27B through Cloudflare Workers AI and Gemini 3.5 Flash-Lite through Google AI for Developers.

Is the video sponsored?

No. Neither the video nor this article is sponsored.

Sources and Update

Model details come from the official Cloudflare and Google model pages; test results come from the video description and our own test notes (as of 14 November 2026). Results are specific to this test and are not a guarantee of any outcome.

Want to apply AI to your own business step by step? Our Skool community is free for 7 days: join the Okyanusi Skool community.

Author: Akın Yılmaz — YouTube: @AKINYILMAZOKYANUSI. Updated: 14 November 2026.

Sources

• https://developers.cloudflare.com/workers-ai/models/qwen3.8-27b/
• https://ai.google.dev/gemini-api/docs/models/gemini-3.5-flash-lite
• https://www.youtube.com/watch?v=5Q1TYvUQB34

Legal and Affiliate Disclosure

This content is for education and general information only; it is not legal, financial, tax or professional advice and does not guarantee income, sales, approval or results. Some links may be affiliate links; I may earn a commission at no extra cost to you.