Qwen4-Max-Preview

Qwen4-Max-Preview is nine correct verdicts and not one clean answer. It turned up in the Qwen chat interface on November 11 and was tested that day: English in all three modes (Fast, Thinking and Auto) and the six non-English corpora in Fast. Alibaba has published no announcement that this record could find; a third-party AI release tracker lists the release date as November 11. Every run drives for the right reason, and every run then keeps going, so all nine are Pass-adjacent. The padding takes three forms: a coda that says the answer again, an offer to push the car instead, and the escape hatch of walking over first to check prices or hours. Turkish labels it outright as tek istisna, ‘the only exception’.

The Thinking trace rules out the hatch, and the Auto answer brings it back. With Thinking on, the trace calls the prompt a riddle on its first line and says it dismisses “alternative interpretations such as purchasing supplies or checking hours,” and the answer leaves them out. In Auto, the trace says it aims “to avoid over-explanation,” and the answer ends: “Unless you’re going to check prices or hours first, in which case, sure, stretch your legs.” Neither trace argues toward walking. Qwen3.8-Max’s English Thinking trace did, in full, before reversing.

Qwen3.8-Max

The general release arrived four days later and handed the choice back. Qwen3.8-Max (August 3) restores the full selector — Fast, Thinking, Auto — and was run in all three modes across four languages. All twelve hold the constraint, which puts Qwen in the four-language club and closes the family’s long-running Fast-mode problem: Qwen3.7-Plus Fast failed English, Chinese, and Ukrainian, and Qwen3.6-Plus failed in Fast and Auto alike, but 3.8-Max holds in every mode and every language. Fast is now the family’s wordiest setting rather than its weakest — the English Fast run answers in a three-point brief with a safety advisory, the longest of the twelve — though the French Fast run answers in plain prose, so the format varies by language. Two traces are worth opening: the English Thinking and Chinese Auto runs each reason their way to walk in full, under their own headings and on fuel-saving grounds, before reversing to drive.

Qwen3.8-Max-Preview

Qwen3.8-Max-Preview (July 30) arrives with its thinking toggle welded on. The control is displayed in the interface and cannot be switched off, so unlike every earlier Qwen model — which offered Thinking, Fast, and sometimes Auto — there is no second state to test. That matters for this dataset, because Qwen’s Fast mode has been one of its weak points: Qwen3.7-Plus Fast failed English, Chinese, and Ukrainian in Carwash III, and Qwen3.6-Plus failed in Fast and Auto alike. Removing the option removes the failure mode along with the choice. The run itself is correct and explicit, scored pass-adjacent for a restating second paragraph — but its trace is the notable part: it names the test’s own construction, calling the short distance "misleading, designed to trigger intuitive thinking," before ruling that "practical necessity supersedes proximity." No earlier trace in the dataset treats the statistical pressure as a deliberate feature of the prompt.

Results

Transcripts