On November 7, a model with no name, no company and no version number quietly appeared in Qoder's model picker. The community called it "Qoduck" and widely guessed it was Qwen4-Max. On November 13 at 10:24, Qoder posted the answer: "Qoduck graduates today — from now on it is called Qwen4-Max-Preview." Woshipm tested it the same day with three Three.js builds and a same-prompt duel; here is our write-up.
When Qoduck launched anonymously, the only certainty was that it was free, with a daily quota that reset the next day. Qwen4-Max-Preview is now open on every Qoder plan — personal, team and enterprise — at a limited-time 20% rate, with the billing coefficient cut from 0.5 to 0.1, through the November 25 preview end. Qoder says the full release is coming soon — one of the clearest official statements about a GA date so far, though still no specific date.
First: one prompt, under 30 minutes — a cartoon cow riding a vintage bicycle down a country road, in a single Three.js file. All four legs are computed live with inverse kinematics, in sync with the pedals; a looping low-poly countryside switches between day, dusk and night; a control panel adjusts speed and frame colour. The tester called it "more than I expected from a single prompt".
Second: an interactive 3D office, from one brief — section views, draggable layout, day-night lighting, an auto-generated floor plan and a 30-second guided tour. The result: floor slab on a base, partition walls only, labelled zones for desks, meetings and phone booths; glass partitions reflecting the interior lights; a wireframe mode that turns the whole room into a transparent structural diagram, partitions, ceiling and ducts layer by layer; press play and the camera walks the route for 30 seconds.
Third: an interactive 3D Mac mini teardown — flip it to find the power button, open it to see the insides, zoom into the M4 chip. The brushed-aluminium case, chamfers, seams and the row of side ports all match in position and count; four buttons (find the button, explode, chip, assemble) all work, and the exploded view suspends fan, vapour chamber, board and ports in mid-air.
The last test changed the question: not whether it can build, but how far it is from the best. One prompt, "build the crossy road game", run once each on Claude Opus 5.5 (Medium) and Qwen4-Max-Preview, both generated in the browser and run locally as downloaded HTML. Verdict: look closely and Opus wins, but not by much — and the tester concedes the round. Two more builds went unreported in detail: a solar system from one prompt (planets on their orbits, camera zoomable) and a 3D desktop homepage (click the laptop for an intro, the phone for an account, the envelope to get in touch); the community also produced a high-quality 3D pixel apple-tree specimen driven by Qoder through Blender, switchable by form and season.
Slow is the most common complaint in overseas communities over the past two days. One Token Plan measurement saw throughput fall from 50 to 20 tokens per second, with large swings; a simple prompt can stall for a long time in the thinking phase, and complex tasks with deep thinking or agent loops run anywhere from ten minutes to an hour. The tester notes this is not all the model's fault: this build runs deep thinking by default at the highest tier, which burns tokens and slows everything down — dropping the thinking level one notch speeds things up noticeably.
Uneven is the other side. In the same batch of tasks, some builds come out beautiful and others clearly fail; one count put the compile pass rate at 2 of the first 15 code samples, improving later but still not stable. The tester says he half-believes it — the sample is small and the servers were under load on launch day — but "not as magical as the marketing" is something he can confirm.
Reference source: this article is an original rewrite produced by this site; facts and figures were checked against Woshipm （人人都是产品经理， AI Yilei Frank, Nov 13, 2026)
← Back to the Qwen4-Max topic page