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

Zhidx Tests the Qwen App: Qwen4-Max's Multimodal Agent Over-Delivers — Consistency Lags

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

Unlike most reviewers, who tested inside coding tools, Zhidx (智东西) tried Qwen4-Max-Preview where ordinary users will meet it: the Qwen App's "Qianwen Go!" mode. The focus was multimodal agency — understanding instructions, searching, generating mixed media and delivering a finished artifact. The verdict in one line: delivery beat expectations; cross-modal consistency did not.

What impressed: a complete agent loop

Instruction understanding and search execution were mostly precise, and the model chained "research — organise — generate" into a single workflow instead of stalling at chat. Multimodal generation quality is clearly up over the previous generation, with 3D simulation and interaction detail standing out. For a model pitched at office productivity and agentic work, that end-to-end delivery matters more than single-turn cleverness.

What broke: subtitles that don't match the picture

The failures were equally concrete: subtitles that mismatch the video they caption — a cross-modal consistency defect — and occasional details that violate basic physics. These bugs only surface when text, image and video must agree inside one deliverable, which is precisely how multimodal agents should be judged: not "can it generate?" but "does everything it generated agree with itself?"

Speed: mid-pack

Zhidx rated responsiveness as moderate. Combined with the ~33 tokens/second anecdotes in other tests (see our Alibaba Cloud developer write-up), Qwen4-Max-Preview sits in "usable, not fast" territory — consistent with a model whose reasoning defaults to its highest tier: depth bought with waiting time.

Our take: if you will use it through the Qwen App for compound tasks that mix search and generation, this test is encouraging — the pipeline completes and the output is presentable. For anything publishable involving captions or image-text correspondence, human consistency-checking is still mandatory.

Reference source: this article is an original rewrite produced by this site; facts and figures were checked against 智东西(2026-11-11)

← Back to the Qwen4-Max topic page