ModelBrief · Qwen4-Max-Preview · Topic Page
Last updated Nov 13, 2026

Qwen4-Max Hands-On: Six Tasks at Maximum Reasoning — Front-End and SVG Shine

Published Nov 13, 2026 · Last updated Nov 13, 2026

With no official benchmarks published, first-wave hands-on tests are the only evidence outsiders have. On launch day, November 11, Chinese reviewer "AI Product Sniper" (techgogogo) set reasoning to its highest tier (xhigh) and ran six tasks spanning graphics, interfaces and complete mini-apps. Here is our restructured take on what those tests actually show.

Task 1: The pelican-on-a-bicycle SVG — a rare clean pass

Drawing a pelican riding a bicycle in SVG is a long-running stress test: body, beak, wheels and pedals must line up spatially or the result falls apart. Qwen4-Max produced what the tester called a rare top-tier result — coherent structure, sane proportions, none of the usual limb chaos. Since SVG is code generation in disguise, this fits the model's coding-and-agents positioning.

Task 2: Compound bow — mechanics right, posture imperfect

A compound bow is hard because the machinery must be real: cams, string path and arrow rest all have to make sense. The pulley, the arrow fletching and the archer figure all came out correct, with none of the common "arrow through the riser" errors. The one flaw noted: the bow's tilt was too shallow, leaving the shooter's stance slightly off. Correctness passed; taste has headroom.

Task 3: A browser operating system — taste in the details

The third task asked for a full desktop OS simulated in a web page — windows, a Dock, a calendar. The tester singled out a small detail most models get wrong: the calendar's date highlight is a rounded rectangle, not the lazy ellipse many models default to. The overall UI and Dock bar were both praised.

Task 4: The space shooter — better than two flagships?

The OS bundle included a space-shooter mini-game, which produced the boldest claim of the set: the tester judged it better than the same-prompt games from Claude Fable 5.2 and Gemini 4 Carbon, Saturn's ring included. Treat this as a single-sample, subjective verdict — no repeated runs, no scoring rubric — but it does place Qwen4-Max in the top tier for one-shot playable games.

Tasks 5–6: Trello board — zero-bug delivery

The final build was a Trello-style kanban board. It shipped functionally complete with no bugs found; the only nit was a slight wobble in neighbouring lists when one list is deleted. "Works, stable, only animation polish left" is exactly the delivery profile that matters for a coding model.

Our take: the six tasks point the same way — Qwen4-Max-Preview excels at structured generation: SVG art, UI components and complete small apps in one shot, with unusual attention to defaults. Its remaining gaps are matters of taste (posture, motion), not correctness. If your workflow is prompt-to-front-end-prototype, this is a strong signal — but it is still one sample, so re-run your own real project before committing.

Reference source: this article is an original rewrite produced by this site; facts and figures were checked against techgogogo(AI产品狙击手)(2026-11-11)

← Back to the Qwen4-Max topic page