How Good Is Qwen4 Max? Hands-On Tests vs Qwen 3.8

How good is Qwen4 Max in real tests?

In one reviewer's hands-on tests of the Qwen4 Max preview, it performed well on long, multi-step work. It built a Blender animation and a full-day street food simulation from single prompts, named a political party in 82 languages without repeating a name, and measured a satellite image by writing its own Python code. It also read a photo more carefully than Qwen 3.8 did. It still thinks a lot, adds text you didn't ask for, and struggled with one complex dish animation.

These are one person's tests, not benchmarks. They are useful because they show how the model behaves on realistic tasks.

What was tested, and by whom?

The tests come from a creator who works with Qwen models, says he is a Qwen ambassador, and says he got early access. He ran the preview in Qwen's studio and through an API in Hermes agent. He says the tests were run on the day of the preview release.

Treat his results the way you would treat a colleague's trial of new software. They are real and specific, but they are one person's prompts, and the model is a preview that may change. Alibaba's own numbers weren't part of this material, so this article doesn't make benchmark claims.

Can it build a full animation from one prompt?

Yes, with caveats. Two tests covered this.

A Blender medical animation. The prompt asked for a face with pronounced forehead lines, a fine needle guided to two points across the brow, small controlled amounts, and the lines softening. It then asked for the lips to be treated in two places for gentle fullness. The model produced a Blender script, and the assets and animation took about an hour to generate. The reviewer found the result impressive.

A rooftop street food stall, start to finish. This was a single HTML, JavaScript and CSS file, not an image generator. The scene ran from before dawn to fireworks at night, with a day counter and a clock. It showed sausages searing on a charcoal grill and eggs frying in a skillet. At midday a doner kebab cone was shaved, with dripping fat, a hint of smoke and a moving sun. The reviewer called it one of the best doner kebab simulations he had seen, while saying a few things could be better. Generation took a little over an hour.

The weak spot was a whole-animal roast dish from northwest Pakistan, cooked on a rotating skewer with rice in the belly. The model struggled with it, and the camera drifted. The reviewer felt it still did the dish some justice.

The lesson: the model can sustain a long, detailed brief. Culturally specific or mechanically complex subjects are where it slips.

How does Qwen4 Max handle 82 languages?

The prompt asked for a political party name in more than 80 languages. Each name had to appeal across left, right and center and across faiths, differ in every language, and feel native to the culture.

According to the reviewer, the model covered all 82 languages with one name each and no repeats. Many names were rooted in local ideas of togetherness. He singled out "Ubuntu" in Zulu and mentioned an Indonesian name based on gotong royong, a term for communal cooperation. For a made-up-sounding case, it used a playful letter-insertion trick.

The miss was instruction following. He asked for names only, and the model added explanations. He also felt it showed "its own personality" compared with earlier Qwen models, which he said follow instructions to the letter. If you need strict formatting, say so firmly and check the output.

Is Qwen4 Max still an overthinker?

Less than Qwen 3.8, but not a light thinker. Reasoning models spend time "thinking" before they answer, and that costs waiting time and, on paid APIs, usually money.

The reviewer's observations:

• Qwen 3.8 was an overthinker. Qwen4 Max improves on that.
• Turning thinking down reduces it, but not by much.
• There is a noticeable gap between low and high thinking. At high, he found performance "way better."
• All his tests used high thinking, which is partly why several ran so long.

If you use the model, expect to choose between speed and quality, and plan for the high setting on hard problems.

How well does it reason about a satellite image?

This was the most demanding test. The prompt gave it a satellite image of southeastern Australia and asked it to:

1. Work out the map scale from town labels.
2. Measure the width of the green coastal strip from Sydney inland.
3. Explain the rain shadow using real lapse rate math.

The reviewer says the model did not guess. It wrote its own Python to find the town labels pixel by pixel. Its first pass confused Sydney and Gosford, and it caught and fixed that. It then measured 12 town pairs and arrived at a scale of about 1 kilometer per pixel.

Its answer was a green band of about 320 kilometers, give or take 50. It judged an early brown patch to be farmland. For the physics, it used tool calls and stated that air cools about 5.5 degrees climbing a 1,000 meter ridge and warms 9.8 degrees descending, so the dry side ends up about 4 degrees warmer. The job took more than 10 minutes. The model named where to draw the red-brown line as its biggest uncertainty, and the image itself was low quality.

Showing its method, correcting itself and flagging uncertainty matters more than the final number. Those habits let you check the work.

Does it read photos better than Qwen 3.8?

In this test, yes. The reviewer gave it a generated image of a girl in a back seat with her seat belt strap under her arm, then asked whether that would be acceptable in the US, Australia, India, Pakistan, Poland, Panama, Chile, Ecuador and Uganda, in practice and not just on paper.

After a long think, the model noticed she was buckled but the strap was under her arm. The reviewer says Qwen 3.8 couldn't infer that. It also separated what the law says from what police enforce. Its verdict: Australia is the only country on the list where this would be a real fine, Poland and Chile carry a small chance, and elsewhere it is practically none. The reviewer was surprised about the US and invited US viewers to confirm.

Check legal answers like these against official sources before acting on them. The useful part is the reasoning pattern, not the verdict.

What this means for you

If you evaluate AI tools at work: Build your own small test set, as the reviewer did. Pick five real tasks from your job: a long document, a messy image, a multilingual request, a data question. Run each on the old and new model with identical prompts.

If you pay per use: Heavy thinking means longer waits and potentially higher costs. Test at both low and high settings on your own tasks.

If you need exact formats: The extra explanations in the naming test are a reminder to specify output format and verify it.

If you want open weights: The reviewer notes that a 27 billion parameter model and a mixture-of-experts version had not been released, contrary to what some social media accounts claim. Don't plan around them until they ship.

A naming caution: community uploads with similar names exist on Hugging Face, such as "Qwen4b-DS-Critic-SFT-Max" quantized files from a third-party uploader. That is a separate project, not Alibaba's Qwen4 Max.

Could you track model tests like these yourself?

A simple scorecard helps if you compare models regularly. With Remy, you can describe a small internal tool in plain English: a form to log each test prompt, the model used, the thinking setting, run time and a 1 to 5 score, plus a page that compares results across models. Remy builds and deploys it as a full-stack app with a database and sign-in, so your team shares one record instead of scattered spreadsheets. Its platform lists nearly 400 AI models, so you could extend the tool later to run prompts directly.