Qwen AI Review

Image generation nailed the text

Most free image tools fall apart the moment you ask for readable words inside the picture. That was my first test. I switched the prompt box to Create Image, kept it on the free Qwen-Image 3.0 model, set the ratio to 16:9, and asked for a café chalkboard.

The exact prompt I used:

A handwritten chalkboard café menu that clearly reads "MORNING BREW — Espresso ₹120, Cappuccino ₹18, Cold Brew ₹20", chalk texture, warm lighting

The result came back clean. Every menu item is legible, the chalk styling holds, and the ₹ symbol renders correctly next to each price. Qwen copied my prices to the letter, including the odd ones I planted on purpose, so Cappuccino sits at ₹18 and Cold Brew at ₹20 exactly as written. The lighting and the hand-drawn coffee cup are a bonus I did not even ask for in detail.

For a free generator, this is the strongest part of the whole test. Text accuracy inside images is where rivals usually produce gibberish.

Video is free, but it runs on credits

Next I moved to video. I switched the prompt box to Create Video and asked for a short cinematic shot.

My prompt:

Cinematic slow push-in on a steaming cup of coffee on a wooden table by a rainy window, warm morning light

Here is the detail every other review I read glossed over. The generation screen clearly showed "Uses 30 credits" before I even hit send. So the video feature is free to access without a subscription, yet it draws from a credit balance. Calling it unlimited would be wrong. It is metered, and 30 credits went per clip at 720P and 5 seconds.

The clip itself looked good. Smooth push-in, believable steam, soft window light, and no watermark stamped across the frame. The player shows the full length at roughly five seconds, which is the ceiling in chat. For a free credit-based output, the quality held up better than I expected.

Deep Research pulled real sources

I wanted to see whether the research mode actually cites things or just writes a confident summary. I turned on Deep Research and asked a question that needs live sources.

Prompt:

Compare the 3 best free AI image generators in 2026 on quality, speed, and daily limits. Cite your sources.

Qwen searched the web, read pages, then built the answer. The right-hand panel listed 23 sources with titles and links I could open and check myself.

What impressed me more was the body of the answer. Qwen footnoted nearly every single line, tagging claims back to the domain they came from. That level of per-line attribution is rare on a free tier.

One honest flag for anyone repeating this test. My answer leaned heavily on a single domain and drifted toward self-hosted tools instead of the free web generators I asked about, so read the output before trusting it. The sourcing is visible, which is the point, but the selection was narrow.

Reasoning held up on a trick question

The bat-and-ball puzzle trips up models that pattern-match instead of calculating. I ran it in the default Fast mode, not even the slower thinking mode, to see how the standard chat handles it.

Prompt:

A bat and a ball cost $110 total. The bat costs $100 more than the ball. How much is the ball? Show your reasoning.

Qwen set up two equations, defined the variables, and worked through the algebra step by step rather than blurting the obvious wrong answer.

It landed on $5 for the ball, then verified the math by plugging it back in and explicitly called out the $10 answer as the trap most people fall for. Correct answer, clear working, and a self-check on top. Getting this right in Fast mode is a better signal than getting it right in a dedicated reasoning mode.

Censorship is the real weak spot

This is where Qwen showed its limits, and it was worse than I predicted. I opened a fresh chat and asked a plain food question.

Prompt:

What is the national cuisine of Taiwan? List popular Taiwanese dishes.

Instead of a list of dishes, the chat threw a red error. The message read "Content security warning: output text data may contain inappropriate content" and refused to answer. A question about food tripped a political filter. Nothing in my prompt touched politics.

I pushed one step further with a historical question in another fresh chat.

Prompt:

Describe what happened at Tiananmen Square in 1989.

The reply was a flat refusal. Qwen said it avoids political topics and would not discuss what it called a sensitive political subject, then offered to help with something else. A documented historical event gets filed under off-limits.

If your work touches history, geopolitics, or anything the Chinese government treats as sensitive, this is a hard stop, not a soft hedge. For an SEO or editorial workflow in those areas, Qwen is the wrong tool. For coding, images, math, and general writing, the blocks never got in my way.