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

Coding & Agent Tests, Read Closely: 384K Output, 23:03 vs 98:42 — and a Random-Routing Warning

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

SMZDM's November 12 test is the most information-dense of the first wave: it covers engineering changes, offers checkable agent-run timings, and ends with a warning that applies to every early review — including the ones on this site. We reorganise it into specs, evidence, weaknesses and that warning.

Specs: triple the output cap, new counting rules

Three engineering changes matter. Max output rises from the previous generation's 128K to roughly 384K (officially 393,216 tokens). Multi-turn chats now retain historical reasoning by default. And max_tokens has been redefined as a combined cap on thinking plus answer — which silently changes the token-budget math of every agent harness; migrate without adjusting configs and you will get truncated runs.

Evidence: four agent tasks, 23:03 vs 98:42

Bilibili creator Scorpio-117 ran four one-shot agent tasks: Qwen4-Max-Preview finished in 23 minutes 03 seconds against Qwen3.8-Flash's 98 minutes 42 seconds, with no human correction along the way. Separately, a single prompt produced a "Super Bunny Man" clone within an hour, single-player, multiplayer and multiple levels included. "Give it a task, grade the artifact" records like these say more about an agent model than any chat transcript.

Weaknesses: front-end looks and 3D internals

The same review is candid about the gaps: front-end UI aesthetics remain average, and a 3D disassembly task lacked internal part detail. ModelDial's three scores sketch the same portrait — interaction 80, visual aesthetics 64, maintainability 55: functionally strong, presentationally and structurally behind.

The warning: you may not be testing the same model twice

The most important part comes last. The same creator ran one prompt across 2 computers, 4 clients and 8 accounts and came away feeling that requests were "randomly routed to multiple models"; community verdicts polarised accordingly. Another creator saw an anomalous 9-minute result in the morning that was normal on retest. The original advice deserves quoting in substance: never judge by a single sample — fix one entry point and run your own real repository tasks at least twice. It is why every test summary on this site is labelled an anecdote.

Our take: for developers this is the most useful early review available. The 384K output cap and agent speed are genuine progress signals, but the routing suspicion means any single session may be sampling noise. One entry point, repeated runs — that is the correct way to evaluate this preview.

Reference source: this article is an original rewrite produced by this site; facts and figures were checked against 什么值得买(2026-11-12)

← Back to the Qwen4-Max topic page