Alibaba launches Qwen4-Max preview; hands-on shows strong multimodal agent delivery but consistency flaws

智东西 · wechat · 2026-11-11

Zhidongxi hands-on with Alibaba's Qwen4-Max preview (Nov 11):

Launches on the Qwen AI platform with up to 90% off via TokenPlan; also live in Qwen Office, Qoder, Qwen app and QwenStudio. Specs and benchmarks undisclosed; the awaited Qwen4-27B hasn't shipped.
Official positioning: stronger world knowledge and perception, upgraded coding and agentic capabilities for real tasks like complex programming, professional office work, frontend design, and software operation.
Highlights: physically plausible pelican-on-bicycle animation with tunable parameters; a train-derailment HTML5 Canvas physics demo with per-car collision volumes and sound; a squirrel-typing GIF with clean subject extraction and background completion; a reconciliation GIF based on a reality-show scene that even lands a Chinese meme.
Issues: multimodal consistency breaks on complex tasks (dialogue assigned to the wrong character, overlapping subtitles), physically implausible details; devs on X found Fable5.2 slightly better, and generation ran at 53 tokens/s.

Verdict: multimodal agent delivery exceeds expectations, but speed, consistency, and detail handling still lag; whether the final version reaches the top tier in coding and agents is the next thing to watch.

Original post →

Alibaba launches Qwen4-Max preview; hands-on shows strong multimodal agent delivery but consistency flaws

More from Models