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

One-Shot 3D Render, Logged Plainly: ~69k Tokens, ~33 Tokens/s, Occasional Intent Drift

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

This short test from the Alibaba Cloud Developer Community is valuable precisely because it is plain: no comparisons, no rankings — just a log of one real one-shot generation, in which Qwen4-Max-Preview wrote a Java + Three.js 3D render in a single pass, consuming roughly 69k tokens.

What "one-shot" proves

3D rendering code is unforgiving about completeness: scene, camera, lights, materials and the render loop are all mandatory and mutually coupled. Producing a runnable result in one pass demonstrates coherence in the "long code, many modules, tight coupling" regime — the same conclusion other tests reach from different directions (the zero-bug Trello board, the one-pass claw machine). One-shot completion is this preview's most consistent strength.

The speed: ~33 tokens/second, again

The author also recorded generation at about 33 tokens per second — matching the figure from Woshipm's claw-machine test. Two independent small samples landing on the same magnitude make "~33 tok/s" the most credible speed reference available, and it is, plainly, not fast. Against a 69k-token consumption, waits are measured in minutes; budget patience for large generations.

The flaw: occasional intent drift

The author notes occasional intent misunderstanding — individual requirements not executed as asked. The sample is tiny and personal; still, alongside Zhidx's cross-modal consistency findings, a pattern emerges: macro-structure is reliable, micro-intent occasionally drifts. Detail fidelity is this generation's homework.

Our take: read this post as a calibration ruler for daily use: dependable one-shot completion, ~33 tok/s speed, occasional intent drift. Together they describe the real texture of working with this preview — and make "check the output against your requirements list, item by item" the highest-value habit you can adopt.

Reference source: this article is an original rewrite produced by this site; facts and figures were checked against 阿里云开发者社区(2026-11-12)

← Back to the Qwen4-Max topic page