AI GUIDEPartner content

Qwen 4 Max and maximum reasoning: what the pelican test showed

In Simon Willison’s test, Qwen 4 Max spent 7 minutes and 23 seconds drawing a pelican in maximum mode. Here’s why longer reasoning doesn’t guarantee a noticeable improvement, and how to compare modes.

Affiliate link: your price stays the same and the project earns a commission.

Qwen 4 Max is a language model, and the reasoning mode in this test is a setting that changes the depth of computation and response time. In a comparison published by developer Simon Willison, the model was asked to draw a pelican on a bicycle. Maximum mode spent 7 minutes and 23 seconds on the task, about 4,2 times as long as the previous level.

A long wait doesn’t prove the mode is useless.

The test shows something more limited: on a simple creative task, the additional computation did not produce an obvious improvement that would justify the wait for every user.

Qwen 4 Max test conditions

The model was asked to create an image of a pelican on a bicycle. Simon Willison compared reasoning levels and noted that maximum mode ran for 7 minutes and 23 seconds. In his estimation, that was about 4,2 times longer than the previous level.

The measurement applies to a specific prompt and a single run. The available information does not confirm that the generation settings, infrastructure, and other conditions were the same in every possible repetition. Without that information, the run time cannot be considered a universal characteristic of Qwen 4 Max.

What did the result look like? The pelican was riding a turquoise bicycle and wearing a red scarf. The experiment’s author noted that the longer reasoning was not accompanied by a noticeable improvement in quality, in his view.

How to assess the drawing

For an image-generation task, it’s helpful to decide in advance what counts as a successful result. In this case, you can check whether the image shows a pelican, whether it is on a bicycle, whether key details are distinguishable, and whether distracting elements have appeared. These criteria make it possible to compare results using the same features, although assessing individual details may still depend on the viewer.

The judgment that something “got better” depends on the criterion chosen. One viewer might prefer the composition or neatness, another the exact fulfillment of the prompt. Without predefined criteria, the comparison remains subjective, even if the run time is measured precisely.

  • Prompt adherence: whether the pelican and bicycle are present.
  • Retention of details: whether the specified or expected elements of the scene are visible.
  • Image coherence: whether there are obvious visual errors that spoil the result.
  • Wait time: how much time passed between submitting the prompt and receiving the finished response.

The bicycle and scarf colors describe the resulting version, but by themselves do not prove that one reasoning level is better than another. Such a conclusion requires comparable results from several modes and a clear way to assess their quality.

Why extra minutes don’t guarantee improvement

A mode with more reasoning increases run time, but duration alone does not measure quality. The model may process a prompt for longer without resolving its ambiguity or changing the result enough for the user to notice a difference.

The pelican task does not test all the model’s abilities: it is not a programming, fact-finding, or complex mathematical problem-solving test. The result of a single visual prompt cannot be generalized to tasks where the model must account for lengthy instructions, check multiple constraints, or fix errors in code.

A lot depends on how the result is assessed: if the differences between two images are minor and maximum mode takes much longer, a user may choose the faster option; if the task is critical and an extra check prevents a significant error, the wait may be worthwhile. The pelican test does not answer this question for other tasks. It shows only that, in this example, the large increase in run time was not accompanied by an obvious improvement in the result for the author.

How to compare reasoning levels on your own tasks

Compare modes using a repeatable task and an unchanged prompt. First, define the success criteria, then run each mode under the same conditions and save the responses along with their wait times. If results vary between runs, repeat the check instead of choosing a mode based on a single successful generation.

  1. Choose a task that can be checked against specific features, such as the presence of specified objects or passing a fixed set of tests.
  2. Keep the prompt and input data unchanged for all levels being compared.
  3. Record the response time and assess each result using the same scale.
  4. Repeat runs if generation randomness noticeably affects the outcome.
  5. Compare the quality improvement with the additional wait and use of the available quota, if the service shows this information.

For a drawing, criteria might include adherence to the description and the absence of noticeable defects. Code is best checked with tests prepared in advance, rather than only by how convincing the response looks.

Choose Max mode based on the results for your own task, not the setting’s name. If a higher level consistently fixes important errors, the extra time may be worthwhile. If the responses barely change, it makes sense to stick with the faster mode.

What the test result means

In Simon Willison’s experiment, maximum reasoning in Qwen 4 Max took 7 minutes and 23 seconds, about 4,2 times as long as the previous level. The resulting pelican was on a turquoise bicycle and wore a red scarf, and the author did not see a noticeable improvement sufficient to justify the wait.

This observation applies to one example, not to all tasks and users.

Practical takeaway: assess the verifiable quality of the result together with the time it took to obtain it, rather than reasoning depth on its own.

Frequently asked questions about Qwen 4 Max reasoning mode

How long did maximum mode take?

In the pelican test, generation took 7 minutes and 23 seconds. This time applies to the specified run.

Did the image get better?

Simon Willison did not note a noticeable improvement that would justify the additional wait. This is his assessment of the result of a specific experiment, not a universal rating of the mode.

Should maximum reasoning always be enabled?

There is no general answer for every task. Compare modes on prompts that reflect your work, and choose the maximum level if it produces a measurable improvement that matters for that particular task.

Compare models before you start

The service sets its plans, limits and model catalog. If they differ from this article, contact us so we can update it and record a new review date.

Browse models

Affiliate link: your price stays the same and the project earns a commission.

Qwen 4 Max reasoning mode maximum reasoning Qwen test pelican on a bicycle AI model comparison

SEO Mind42 editorial team

We explore SEO and neural networks in practice: test services on our own projects, verify prices and limits against primary sources, and share things you can put to use the same day.

📚 Reference guide to SEO and AI 🔄 Materials are updated 🕐 Updated: 19 November 2026

Related reading

All in this section →