Hands-on tests of Qwen4 Max covering animation generation, 82-language naming, satellite image reasoning, and visual safety judgment.
What is Qwen4 Max and what makes it different? #
Qwen4 Max is Alibaba’s latest large language model, released in preview form and tested here through Qwen Chat’s studio interface and via API. The headline difference from its predecessor, Qwen 3, is less overthinking and a stronger tool-use loop: in hands-on testing, the model wrote and ran its own Python to verify claims instead of guessing, and it showed a more distinct “personality” in open-ended tasks compared to Qwen 3’s stricter instruction-following. A separate, larger Qwen release (referred to informally as a 480B or mixture-of-experts model) has not shipped yet, despite claims circulating online.
TL;DR #
- Single-prompt animation generation produced a Blender-based medical simulation (a Botox-style procedure) and a full day-night rooftop food stall scene in HTML, CSS, and JavaScript, each built from one detailed prompt and roughly an hour of generation time.
- Multilingual naming across 82 languages came back with no repeated names, culturally grounded choices (like Ubuntu in Zulu and Gotong Royong in Indonesian), and consistent quality even for a deliberately tricky “gibberish” language request.
- Satellite image reasoning went beyond description: the model derived a map scale from town labels, caught and corrected its own mix-up between two towns, and used real lapse-rate physics (about 5.5°C cooling per 1,000m ascent, 9.8°C warming on descent) to explain a rain shadow effect.
- Visual safety and legal reasoning correctly identified that a child’s seatbelt was buckled but worn under the arm, then separated written law from real-world enforcement across nine countries, flagging Australia as the one place a fine is actually likely.
- Thinking depth is adjustable but even on “low,” the model still reasons substantially more than earlier Qwen versions, and performance noticeably improves when thinking is set to “high.”
- No hallucination was observed in the harder reasoning tasks. The model stated its own uncertainty (for example, flagging exactly where it was least confident on the rain shadow boundary) rather than inventing a confident but wrong answer.
Built like a system. Not vibe-coded.
Remy manages the project — every layer architected, not stitched together at the last second.
How did Qwen4 Max perform on creative and animation tasks? #
The first test asked the model to simulate a medical cosmetic procedure (a Botox-style treatment: softening forehead lines, adding lip fullness) rendered in Blender through a single detailed prompt and an accompanying script. The output matched the prompt’s specific requests, including precise injection points and visible before-and-after changes, generated with assets in about an hour.
A second, more ambitious test asked for a continuous, one-day simulation of a rooftop street food stall, from pre-dawn to fireworks at night, built entirely in HTML, CSS, and JavaScript rather than through image or video generation. The scene included a day/time counter, a waking city, a charcoal grill searing sausages and eggs, a donor kebab stand with visible fat drippings and smoke, and a late addition of a regionally specific dish (a whole roasted animal stuffed with rice, cooked on a rotating skewer). That last dish was where the model visibly struggled, particularly when the camera angle shifted, though the broader simulation (lighting changes, sun position, nighttime fireworks) held up well. The test suggests Qwen4 Max can translate a long, structured natural-language brief into functioning front-end code that produces a coherent animated scene, not just a static visual.
Can Qwen4 Max handle multilingual and culturally specific naming? #
One test asked the model to invent a political party name in over 80 languages, with the constraint that each name had to feel native to its culture, avoid repeats, and appeal broadly across the political spectrum. Qwen4 Max returned a unique name for each of the 82 languages requested, several of which drew on real local concepts of unity or togetherness: Ubuntu in Zulu, Gotong Royong in Indonesian, and similar culturally rooted terms in other languages. It also handled a deliberately awkward request involving invented or “gibberish” language rules with a workable, playful approach rather than failing outright.
The main deviation from the prompt was minor: the model added brief explanations alongside some names when it was asked only for the names themselves. That aside, the range and cultural specificity of the output, covering a wide span of languages without obvious repetition or generic filler, points to meaningfully strong multilingual grounding rather than surface-level translation.
How well does Qwen4 Max reason over scientific images? #
This was the most demanding test: a satellite image of southeastern Australia, with instructions to calculate the map’s scale from town labels, measure the width of a green coastal strip between two cities, and explain the region’s rain shadow effect using actual atmospheric physics.
Without shadow IT.
Every app, approved
✓
✓
✓
✓
✓
Qwen4 Max did not simply describe the image. It wrote its own Python code to locate town labels pixel by pixel, initially confused two nearby towns, caught the error itself, and corrected it. It then measured distances across 12 town pairs to establish a scale of roughly 1 kilometer per pixel. From there it estimated the green coastal band at about 320 kilometers (with a stated margin of error of roughly 50 km), identified a likely farmland boundary, and explained the rain shadow using standard lapse rate figures (approximately 5.5°C of cooling per 1,000 meters of ascent and 9.8°C of warming on the descending side), concluding the dry side would run about 4°C warmer. The full process took over 10 minutes, and the model explicitly flagged where it was least confident, the exact placement of the vegetation boundary, rather than presenting an estimate as fact.
That combination (self-correction, tool use, physics grounded in real constants, and stated uncertainty) is a meaningfully higher bar than typical image captioning or description, and it was tested on a source image of middling quality.
Is Qwen4 Max good at nuanced legal and visual reasoning? #
The final test combined image understanding with cross-jurisdictional legal judgment. The model was shown an image of a child in a car’s back seat wearing a seatbelt incorrectly (routed under the arm rather than across the shoulder) and asked whether this would actually result in a problem in nine different countries: the US, Australia, India, Pakistan, Poland, Panama, Chile, an unspecified “Eastor,” and Uganda.
Qwen4 Max first correctly identified the specific visual detail that the strap was under the arm rather than improperly removed, something the previous Qwen 3 model reportedly failed to catch. It then distinguished between what the law technically says and how strictly it’s enforced in practice, concluding that Australia was the only country on the list where a real fine was likely, that Poland and Chile carried a small chance of enforcement, and that enforcement elsewhere was minimal to nonexistent, even where the behavior is technically illegal (including the US). It also correctly noted that enforcement in such cases typically targets the driver, not the child.
This type of task, reading a specific visual detail accurately and then reasoning about real-world norms rather than just citing statute, is a harder and more practically useful skill than simple image labeling, and it’s where the gap between model generations showed up most clearly.
Frequently Asked Questions #
What is Qwen4 Max?
Qwen4 Max is a large language model from Alibaba, currently available as a preview through Qwen Chat and via API. It’s distinct from a separate, larger mixture-of-experts Qwen model that has been rumored but not yet released.
How is Qwen4 Max different from Qwen 3?
In hands-on testing, Qwen4 Max showed less excessive “overthinking” than Qwen 3, stronger self-correction during multi-step reasoning (such as catching its own labeling errors on a map), and better accuracy on fine-grained visual details, like correctly identifying an improperly worn seatbelt strap where Qwen 3 reportedly missed it.
Does Qwen4 Max support image and visual reasoning?
Yes. It was tested on a satellite image requiring scale calculation and rain shadow physics, and on a photo requiring correct identification of seatbelt placement combined with country-specific legal reasoning. In both cases it combined visual reading with calculation or research rather than offering a surface-level description.
Can Qwen4 Max generate animations or visual simulations?
It can generate code-driven simulations from a single prompt, including a Blender-based medical procedure animation and a full day-night HTML/CSS/JavaScript scene of a street food stall, though complex or unusual subjects (like a specific regional dish) can still trip it up.
Is the full, larger Qwen model (sometimes called Qwen 480B or MoE) available yet?
regulatory compliance and security.
No. As of this testing, only Qwen4 Max preview had been released. A larger mixture-of-experts version has been anticipated and discussed online but had not been officially released.