Qwen 4 Max: an honest review and practical test results

Qwen 4 Max is a language model in a pre-release version. The author tested it on generating 3D animation, a browser game, and an interactive website. At the time of testing, the model was said to support a context of up to 1 million tokens, but the author could not find available information about its benchmarks or cost.

The experiment’s main conclusion is restrained: Qwen 4 Max can assemble working prototypes, but, in the author’s subjective assessment, it has not yet established itself as a flagship model.

In a comparison of impressions, DeepSeek performed more convincingly, especially in the interactive aquarium scenario.

Qwen 4 Max status at the time of testing

During testing, Qwen 4 Max was in preview, meaning it was in a pre-release state. The results apply to a specific version and do not describe the model after possible updates. The author assumes that quality may improve after the full release, but the available materials contain no confirmed plan for such an improvement.

The stated context was 1 million tokens. A large context by itself does not prove high programming quality, resilience on long tasks, or the ability to correct errors carefully, since these properties are checked through separate tests, and official benchmark results were absent from the experiment under review.

Access depended on the platform. On Alibaba Cloud, the author purchased access to Qwen 4 Max through a subscription for 3 dollars. It was not possible to purchase the model through Qwen Cloud: the service requested documents. This experience relates to the author’s specific access conditions and does not confirm universal model availability or a single rate for all users.

Experiment modes and conditions

The tests used the highest available reasoning level, X High, and the Alibaba Token Plan. This mode shows what the model can do with extended reasoning time and resources, but at the same time makes the results less representative of fast and inexpensive workflows.

The author tested not only text responses, but also finished interactive results created through a browser-based workflow. The evaluation covered the scene’s visual integrity, object behavior, mechanics, camera stability, and the interface’s suitability for further use.

The author could not determine the actual monetary expenses for individual generations. Execution time was recorded for several tasks, but it cannot automatically be considered the model’s own speed: duration is affected by the agent process, browser actions, service load, repeated operations, and the amount of required corrections.

Flower animation generation

The flower animation took 45 minutes 37 seconds to create. The seed opened correctly, leaves and a bud appeared, and the background and main movement looked functional. One separate drop hung in place, and the bud opened sideways instead of fully opening in all directions.

For a demonstration prototype, the result looks acceptable, although the errors are noticeable without special inspection. The model completed the overall compositional task, but did not bring individual elements that viewers notice immediately to a polished state.

Orc scene

The orc scene took 32 minutes 13 seconds to generate. Qwen 4 Max created the environment, target, rocks, and campfire, and also added animations for lightning, the character’s swaying, a fireball, and a totem.

The work on the character itself was weaker. According to the author’s observation, the orc looked more like a troll, the fangs and staff were rendered incorrectly, and the model interpreted some details differently from what the description required. The animation effects looked better than the basic 3D model.

This episode shows the difference between having many objects and executing them accurately. A scene may seem rich, but a key character’s mismatch with the original task reduces the result’s value for a game, presentation, or commercial prototype.

Aquarium: the most revealing test

The aquarium took 53 minutes 27 seconds to create. The model added rocks, sand, a crab, a snail, fish, a castle, a bubbler with bubbles, micro-particles, and plants. From a static angle, the scene looked rich and demonstrated a wide range of objects.

Problems appeared after the scene was rotated. The crab barely moved and merged with the snail, the snail’s shell had a strange shape, its movement looked incorrect, and when the scene rotated, its head separated from its body. The fish swam unnaturally fast and moved every part of their bodies.

The environment’s geometry also failed the test from different angles: the castle separated, a separate tower shifted, the walls were built incorrectly, and parts of the plants and other objects “floated.” Some elements looked normal only in particular camera positions, indicating weak spatial consistency in the result.

The author assumed that, based on the interface and the process behavior, the model might have used Playwright and taken screenshots during generation. This is an assumption, not an established technical fact. Even if such a workflow was used, it did not eliminate the noticeable errors.

Based on the aquarium scenario, the author subjectively rated Qwen 4 Max below DeepSeek. The comparison cannot be considered a standardized benchmark: the materials do not contain a full description of identical prompts, the number of repetitions, evaluation criteria, or control conditions.

Browser game

The game took 1 hour 21 minute to create, after which the author asked the server to launch the result. The main mechanics, including the enemies, worked, and the playthrough ended after the character reached the light.

The execution quality remained at the prototype level. The graphics were simple, and the character continued sliding and running after the key was released, making it easy to fall. The camera could move downward or too slowly, and this defect interfered with control more than the lack of visual detail.

The author also noted similarities between the level and mechanics and previously seen generations by Fable and Gemini, but did not confirm the origin of this similarity. This observation cannot be turned into a claim of copying: without analyzing the source code, prompts, and generation history, the cause of the match is unknown.

Studio website

The studio website took 43 minutes to create. The page featured sounds, micro-particles, and a 3D object that the user could rotate and recolor. According to the author, the site was generally responsive and worked acceptably.

Visually, the website looked unusual and did not create an unpleasant impression thanks to the 3D element. At the same time, this object distracted too much from the main information, while the sound design raised doubts. Qwen 4 Max handled creating an impressive demonstration better than prioritizing elements in the interface.

Strengths and weaknesses according to the test results

Qwen 4 Max and DeepSeek: what can be said accurately

The author was left with mixed impressions of Qwen 4 Max: the model seemed decent to him, but in his opinion DeepSeek worked better. The difference was especially noticeable in the aquarium scenario, where rotation testing revealed errors in the geometry and object behavior.

These observations cannot be used to derive a universal ranking of the two models. The experiment covered a limited number of tasks, did not include an independent quality measurement, and did not present a complete replication methodology. The more accurate conclusion is: in the scenarios described by the author, DeepSeek left a stronger impression, while Qwen 4 Max did not demonstrate clear superiority.

Final verdict

At the time of testing, Qwen 4 Max looked like a capable but raw tool for generating prototypes. It created rich 3D scenes, an interactive website, and a game, but lengthy execution did not guarantee a polished result: key errors were discovered after checking the animations, controls, and different camera angles.

In the author’s subjective assessment, the model did not match the level of a flagship model and was inferior even to some mid-range solutions. The author also believes that more interesting models exist for similar tasks. This opinion is based on the experiment conducted, not on an official comparative study.

Should you choose Qwen 4 Max for development? For a one-off visual prototype, the model may prove useful if the developer is prepared to inspect the scene manually and correct the code. For a project where stable geometry, precise controls, correct animation, and predictable expenses matter, the result of this test alone is insufficient.

The main caveat concerns its preview status. Quality may indeed change after release, but this is currently the author’s assumption, not a confirmed promise from the developer. It is reasonable to make a final assessment of Qwen 4 Max only after a stable version, official benchmarks, clear access conditions, and repeated tests on the same set of tasks become available.

Frequently asked questions

Can Qwen 4 Max already be considered a flagship model?

Based on the described test, the author does not think so. The model created working prototypes, but made noticeable errors in details, geometry, animation, and controls.

What context length was stated for Qwen 4 Max?

The test materials state a value of 1 million tokens. This is a stated context specification, not proof of the model’s quality on tasks of that size.

How much did the test cost?

The author purchased access on Alibaba Cloud for 3 dollars through a subscription, but could not determine the actual expenses for individual generations. Therefore, the exact cost of the experiment cannot be stated based on the available data.

What should you choose for similar tasks: Qwen 4 Max or DeepSeek?

In the described experiment, the author preferred DeepSeek, especially after testing the aquarium. For a final decision, it is worth running identical prompts yourself and checking not only the attractive first frame, but also object behavior, controls, responsiveness, and error correction.