ModelBrief · Qwen4-Max-Preview · Topic Page
Last updated Nov 13, 2026
IN PREVIEW · RETIRES NOV 25

Qwen4-Max-Preview: The Topic Page

Alibaba's first Qwen4-generation flagship opened a two-week preview on November 11, 2026: reachable only through a Token Plan subscription and Alibaba's own apps, with no public per-token API price and no benchmark scores. This page collects the official specs, the five ways to try it, Token Plan pricing, and the first hands-on reviews in English and Chinese — with unverified claims clearly marked.

Countdown to preview endThe model retires after Nov 25, 2026, 24:00 Beijing time
--days
--hrs
--min
--sec
1M tokensContext window
393,216Max output tokens (shared with reasoning)
14 daysPreview window: Nov 11 → Nov 25
From $6/moToken Plan (new users $3 first month)
NoneBenchmarks: none published yet
Not openNo weights, no model card

01Key facts

What is confirmed: this is a time-boxed preview, not a general release.

Launch
2026-11-11 10:00 Beijing time (02:01 UTC); the Qwen team posted "Qwen4 is coming" on X
Preview window
2026-11-11 10:00 → 2026-11-25 24:00 Beijing time. The model is then retired; no replacement named
Model ID
qwen4-max-preview (alias qwen4.0-max-preview)
Context
1,000,000 tokens (per Alibaba Cloud Model Studio docs; flagged as unconfirmed in some Nov 11 coverage, confirmed by later docs)
Max output
393,216 tokens, shared between internal reasoning and the final answer. Qwen Code guide default output limit: 32,768 tokens. (Previous gen Qwen3.8-Max: 131K output spec.)
Reasoning
On by default; levels low / medium / xhigh (xhigh default)
Input / output
Input: text, image, video (in chat also documents and audio). Output: text
Built for
In Alibaba's words: complex coding, front-end design, office productivity, agentic tasks
Access
API via Token Plan subscription only; also in Qwen Studio chat, the Qwen app, QwenWork and Qoder. No pay-as-you-go API — a standard Qwen Cloud API key cannot reach it
Price
No per-token price published. Token Plan from $6/month (new users $3 first month)
Benchmarks
NONE — official or independent (Artificial Analysis and LMArena had no score as of Nov 12)
Weights / license
None published. Not on Hugging Face. Not open source
The Qwen4 family
Announced 2026-09-22 at the Apsara conference in Hangzhou, in four tiers: Max, Plus, Flash, 27B (local deployment). Only Max has shipped, as a preview; the other three have no dates

02Timeline

From the Apsara announcement to retirement day — plus one historical precedent.

  • 2026-09-22

    Qwen4 family announced at the Apsara conference; "currently in training"; four tiers: Max, Plus, Flash, 27B.

  • 2026-11-02

    Alibaba API compatibility notice: max_tokens now means max_completion_tokens (reasoning + answer share the cap).

  • 2026-11-06

    "The becoming begins" teaser.

  • 2026-11-07

    Qoder starts testing the anonymous model "Qoduck" for personal users; the community widely believes it is Qwen4-Max. Alibaba has never confirmed this.

  • 2026-11-11 02:01 UTC

    "Qwen4 is coming" post; Qwen4-Max-Preview opens.

  • 2026-11-11/12

    Coverage wave: Apidog, Quantslant, AI First Look, TechTimes, Neurohive and Chinese outlets.

  • 2026-11-25 24:00 Beijing

    Preview ends; the model is retired.

  • Precedent (NOT a promise)

    Qwen3.8-Max went preview Jul 19 → GA Aug 3 → open weights Aug 12, 2026 — about three weeks. Whether Qwen4 follows the same path is unknown.

03Official specs

Only what Alibaba has published. Anything undisclosed is stated as such — nothing is filled in by guesswork.

Status
Preview (Nov 11–25, 2026)
Model ID
qwen4-max-preview
Context window
1M tokens
Max output
393,216 tokens (shared with thinking; 32,768 default in Qwen Code)
Reasoning
On by default (low / medium / xhigh)
Input
Text, image, video (+ documents / audio in chat)
Output
Text
API
Token Plan only, Singapore region
Benchmarks
None
Model card
None
Parameters
Undisclosed
Weights
None released
Price per token
Not published

04How to try it: 5 routes

Chat is the fastest route; for developers there is exactly one: Token Plan.

1Qwen Studio (web chat)

The simplest route. Open chat.qwen.ai and pick "Qwen4-Max-Preview" in the model picker. An account is required; there is a free chat quota (one tester ran a full game test for free).

2Qwen App

Early access: update to the latest version and switch the model in the picker. Free quota for all users.

3QwenWork

Alibaba's office agent app; members get limited-time free access to the preview.

4Qoder

Alibaba's coding environment; available on all plans, including Teams and Enterprise.

5Token Plan API (developers)

QwenCloud account, Singapore region only; keys start with sb-sp-. OpenAI- and Anthropic-compatible; works with Claude Code, Qwen Code, Cursor, Cline, Codex, Kilo CLI, OpenCode and OpenClaw. QwenCloud's Claude Code setup maps Qwen4-Max to the Opus slot.

Base URL: https://token-plan.maas.qwencloudapi.com/compatible-mode/v1 (also an ap-southeast-1 variant) Model: qwen4-max-preview
Token Plan limits (official terms): personal interactive development ONLY, through designated tools. No production automation, batch scripts or background scheduled tasks; non-refundable; no account sharing; no free trial.

05Token Plan pricing, decoded

There is no pay-per-token price for Qwen4-Max-Preview — access is sold as monthly credit bundles. Prices checked Nov 11, 2026; the checkout page governs.

PlanMonthlyNew users (until Nov 30)Credits/moConcurrent agents
Lite$6$320,0001–2
Essential$10$545,0002–3
Standard$18—80,0003–4
Pro$43—200,0006–8

Extra bundle: $15 per 20,000 credits. Note: Apidog lists different list prices ($8 / $16 / $25 / $51) — trust the official checkout page.

Preview discount: during the preview, Qwen4-Max burns credits at as low as 20% of the standard rate — and 8% during Beijing night hours 22:00–08:00 (14:00–00:00 UTC). Chinese coverage describes it as "as low as 10% off-list." However, the standard credit rate per request is NOT published: your real cost is only visible in the console after use. That is the biggest transparency gap in the pricing.

Official pages: QwenCloud Token Plan · Qianwen AI Platform Token Plan

Order Token Plan: China region — Order Token Plan to try Qwen4 · Global — Order Token Plan to try Qwen4

06Early hands-on reviews

There are no formal benchmarks, so everything below is first-wave hands-on testing from Nov 11–12: small samples, varied methods — treat each as an anecdote, not a measurement. Every entry carries a verdict badge; "Read the full article" opens an in-depth rewrite hosted on this site, with the original source linked at the end of each article.

English sources

AI First Look: 8 hands-on tests collected

Verdict: mixed — 4 positive, 4 mixed, 0 negative

Source: AI First Look, Nov 12 (tests/comments collected Nov 7–12)

Positives: a 3D solar system and a neon puzzle game built in Qoder; one coding task matched Qwen3.8-Max (20/20 tests passed) in ~2.5 minutes instead of 17. Mixed: a T-Rex game test finished faster than Claude Fable 5.2 and Gemini 4 Carbon, but the tester still rated Fable 5.2 best overall; one Hacker News user found it slightly slower than GPT-6.1 Terra.

Read the full article on this site → (original rewrite, hosted here)

StealthMark: first independent measurement

Verdict: visual up, general down

Source: Stealth Models, Nov 11–12

Qwen4-Max-Preview scores higher than Qwen3.8-Max in visual intelligence, lower in general intelligence (two games fell most). Editorial estimates only; small sample.

Read the full article on this site → (original rewrite, hosted here)

AIHubMix: first per-token API + an empty-reply bug

Verdict: a third-party price reference

Source: AIHubMix (Medium), Nov 11

The first platform offering it as a per-token API: $0.353 per 1M input tokens, $1.059 per 1M output tokens (third-party pricing, not Alibaba's). It also found a bug: a 300-token reply with nothing in it.

Read the full article on this site → (original rewrite, hosted here)

Neurohive / Qoder: 11 user builds in 48 hours

Verdict: strong at building complete apps

Source: Neurohive, Nov 11–12

Qoder collected 11 user builds from the first 48 hours: a voxel White House with 68 walkable rooms, a floating palace in a single HTML file, a QR-code Chrome extension shipped the same day. Qoder's own comparison put Qoduck at 40 minutes vs DeepSeek V4.1 Flash at 70 minutes on one task (vendor marketing, not independent). X reactions were mostly positive; one tester preferred it to Claude Opus 5.5 on the same prompt (one-off, no methodology). 53AI was more cautious: fast, writes simple code cleanly.

Read the full article on this site → (original rewrite, hosted here)

Chinese sources (summarized in English)

techgogogo: 6-task test at highest reasoning

Verdict: excellent SVG & front-end generation

Source: techgogogo / "AI Product Sniper", Nov 11

(a) The pelican-riding-a-bicycle SVG was called a "rare top-tier result"; (b) compound-bow SVG good — pulley, arrow fletching and figure all correct, though the bow tilt could be bigger; (c) a browser OS with good UI — the date highlight is a rounded rectangle rather than an ugly ellipse, the Dock bar is good, the space-shooter was judged "better than Fable 5.2 and Gemini Carbon," and Saturn's ring is beautiful; (d) a Trello kanban with zero bugs (minor: deleting a list makes right-side lists wobble slightly).

Read the full article on this site → (original rewrite, hosted here)

Zhidx (智东西): tested in the Qwen App

Verdict: delivery beat expectations; consistency issues

Source: Zhidx, Nov 11–12 (tested in the Qwen App's "Qianwen Go!")

Multimodal agent delivery exceeded expectations; instruction understanding and search execution were mostly precise; multimodal generation quality clearly improved; 3D simulation and interaction detail are high; speed is moderate. Problems: multimodal consistency issues (subtitles mismatching video) and some physics-defying details.

Read the full article on this site → (original rewrite, hosted here)

Woshipm (人人都是产品经理): the claw-machine challenge

Verdict: solid delivery, but slow

Source: Woshipm, Nov 12

The claw-machine build took over an hour; output ran at ~33 tokens/s (slow); the final delivery was solid — mechanical force feedback, rope sway, doll collision and scoring all present. Judged "on par with Kimi K3.1."

Read the full article on this site → (original rewrite, hosted here)

SMZDM (什么值得买): engineering view + agent runs (with a big caveat)

Verdict: don't judge by one sample

Source: SMZDM, Nov 12

Engineering changes: max output 128K → 384K; multi-turn chats retain history reasoning by default; max_tokens now caps thinking + answer together (this changes harness token budgets). Agent evidence: Bilibili creator Scorpio-117 ran 4 one-shot agent tasks in 23:03 vs Qwen3.8-Flash's 98:42, with no human correction; one prompt recreated a "Super Bunny Man" clone in 1 hour (single + multiplayer, multiple levels). Weaknesses: front-end UI aesthetics still average; 3D disassembly lacks internal part detail. ModelDial scores: interaction 80, visual aesthetics 64, maintainability 55.

Big caveat: the same Bilibili creator tested one prompt across 2 computers / 4 clients / 8 accounts and felt "multiple models were randomly routed" — community verdicts are polarized; another creator got an anomalous 9-minute result in the morning that was normal on retest. The quoted advice: don't judge by one sample; fix one entry point and run your own real repo tasks twice.

Read the full article on this site → (original rewrite, hosted here)

Alibaba Cloud Developer Community: one-shot 3D render

Verdict: anecdotal record

Source: Alibaba Cloud Developer Community, Nov 12

A one-shot Java + Three.js 3D render using ~69k tokens; generation speed ~33 tokens/s; occasional intent misunderstanding (small-sample personal experience).

Read the full article on this site → (original rewrite, hosted here)

SMZDM (office piece): long-horizon execution is the gain

Verdict: usable for office work; slow thinking

Source: SMZDM, Nov 12

Office verdict: the gain is long-horizon task execution, not chattier chat. Tested working: summarizing group chats, assigning tasks, creating todos, deadline reminders. Shortcomings in the same feedback: slow thinking, front-end UI generation still average; back-end/office scenarios "acceptable."

Read the full article on this site → (original rewrite, hosted here)

06BArticles: in-depth reads on this site

All 12 pieces below are original rewrites produced for this site — restructured, re-argued and checked fact-by-fact against the original reviews. Read them here; each ends with a link to its source for attribution.

Qwen4-Max Hands-On: Six Tasks at Maximum Reasoning — Front-End and SVG ShineOur in-depth write-up of techgogogo's six-task test of Qwen4-Max-Preview: pelican SVG, compound bow, a browser OS, a space shooter and a Trello board, task by task. Zhidx Tests the Qwen App: Qwen4-Max's Multimodal Agent Over-Delivers — Consistency LagsOur write-up of Zhidx's hands-on in the Qwen App: multimodal agent delivery beat expectations and search execution was precise, but subtitles mismatched video and some details defied physics. The Claw-Machine Challenge: Over an Hour of Grinding, a Solid Build, and a Speed ProblemOur write-up of Woshipm's claw-machine test: 1+ hour, ~33 tokens/s output, but force feedback, rope sway, collisions and scoring all delivered; judged on par with Kimi K3.1. Coding & Agent Tests, Read Closely: 384K Output, 23:03 vs 98:42 — and a Random-Routing WarningOur deep read of SMZDM's engineering test: max output up to 384K, Scorpio-117's four agent tasks at 23:03 vs 98:42, a one-hour Super Bunny Man clone — plus inconsistent results across accounts suggesting random routing. Office Tasks Tested: Qwen4-Max's Gain Is Long-Horizon Execution, Not Chattier ChatOur write-up of SMZDM's office test: group-chat summaries, task assignment, todos and deadline reminders all worked; slow thinking and average front-end output are the costs. Eight Hands-On Tests, Reviewed: 4 Positive, 4 Mixed — Qwen4-Max's First Report CardOur review of AI First Look's roundup of 8 early tests: a 3D solar system and puzzle game in Qoder, a coding task passed 20/20 nearly 7x faster — while other testers still rated Fable 5.2 best overall. Reading the First Independent Scores: Visual Intelligence Up, General Intelligence DownOur analysis of StealthMark's first independent measurement: Qwen4-Max-Preview beats Qwen3.8-Max on visual intelligence but trails on general intelligence — and why these numbers are indicative only. A Third-Party Per-Token Price Appears: $0.353/$1.059 per 1M — Plus a 300-Token Empty ReplyOur write-up of AIHubMix's day-one findings: the first per-token listing for Qwen4-Max-Preview at $0.353 in / $1.059 out per 1M tokens, and a bug where a 300-token reply came back empty. 48 Hours After Launch: What Did Qoder Users Actually Build with Qwen4-Max?Our rundown of 11 user builds from the preview's first 48 hours in Qoder: a 68-room voxel White House, a single-file floating palace, a same-day Chrome extension — with the 40-vs-70-minute claim flagged as vendor marketing. One-Shot 3D Render, Logged Plainly: ~69k Tokens, ~33 Tokens/s, Occasional Intent DriftOur write-up of an Alibaba Cloud developer's one-shot test: a Java + Three.js 3D render generated in a single pass using ~69k tokens at ~33 tokens/s, with occasional intent misunderstanding. Qwen4-Max vs Qwen3.8-Max: Should You Switch? A Point-by-Point ComparisonOur head-to-head of Qwen4-Max-Preview and Qwen3.8-Max: status, API access, 393K vs 131K max output, pricing and AA index — with a clear verdict: production stays on 3.8, evaluation goes to 4-Max. Cross-Vendor: Qwen4-Max vs Claude, Gemini and GPT — Why It Cannot Be Ranked YetOur cross-vendor comparison: Claude Fable 5.2 (AA 65), Gemini 4 Carbon (60), GPT-6 Astra (57) and Sol (55) all have scores; Qwen4-Max is not listed — and without a score there is no honest ranking.

07Comparisons

vs Qwen3.8-Max (predecessor)

DimensionQwen4-Max-PreviewQwen3.8-Max
StatusPreview (Nov 11–25)GA since Aug 2026
APIToken Plan subscription onlyPublic pay-per-token: $2 in / $6 out per 1M
Context1M1M
Max output393K (32,768 default in Qwen Code)131K
BenchmarksNoneVendor table published
Artificial Analysis indexNot listed45 (Sep 27, 2026)

Apidog's verdict: pick Qwen3.8 for production; test Qwen4-Max on your own work.

Read the deep dive: Qwen4-Max vs Qwen3.8-Max →

vs rival flagships (Artificial Analysis index, via Quantslant, Nov 11)

ModelAPI price $/1M in/outAA indexContext
Qwen4-Max-PreviewToken Plan subscription onlyNot listed1M
Claude Fable 5.28 / 4065 (max)1M
Gemini 4 Carbon2 / 10 (until Nov 30)60 (high)1M
GPT-6 Astra10 / 5057 (max)1.05M
GPT-6.1 Sol2 / 1055 (max)1.05M
Qwen3.8 Max2 / 645 (Sep 27)1M
Honest note: until Qwen4-Max has numbers of its own, it cannot honestly be ranked above or below anyone in this table. Any "ranking" circulating right now is not a measurement.

Read the cross-vendor comparison: Qwen4-Max vs Claude / Gemini / GPT →

08What's still missing (as of Nov 12)

  • Benchmarks (vendor or independent)
  • Model card
  • Parameter count
  • Per-token price
  • Open weights / license
  • GA date
  • Dates for Qwen4-Plus / Flash / 27B
  • Speed/latency measurements (beyond ~33 tok/s anecdotes)
  • Safety evaluation

09Rumor check

The following claims are circulating but unverified. Do not cite them as fact.

UNVERIFIED "Qwen4-Max scores 53 on the Artificial Analysis index"

A chart circulating on X/Reddit. Checks: the score is absent from AA's own leaderboard and posts; Stealth Models could not verify it; the chart's own legend stripes models without a public API, yet this bar is solid; the arrow is drawn by hand.

UNVERIFIED "Qwen4-Coder scores 82% on SWE-bench Verified"

The page that published it calls the figure illustrative and cites no primary source; Alibaba has announced NO Qwen4-Coder (the current coding models are the Qwen3-Coder line).

UNVERIFIED Architecture claims (GDN + QSA hybrid attention, gated residuals, N-gram embeddings, Muon optimizer)

Sourced to a single Chinese aggregator. Alibaba has not described the architecture (as TechTimes also notes).

UNVERIFIED Social posts claiming it "beats GPT-6 and Claude Fable 5"

There is no confirmation or data behind them.

10FAQ

Is Qwen4-Max free?

Chat in Qwen Studio / the Qwen App is free with an account (limited quota). API access requires Token Plan ($3–6/month to start). Token Plan has NO free trial.

Is it open source?

No. No weights and no license have been published.

What happens on Nov 25?

The preview model is retired. No replacement has been named and the GA date is unknown. (Precedent: Qwen3.8-Max went preview → GA → open weights within ~3 weeks, but that is not a promise.)

Can I use it in production?

No. Token Plan is for personal interactive development only — no automation, batch jobs or scheduled tasks.

Is it better than Claude / GPT / Gemini?

Unknown — there are no independent benchmarks yet. Early hands-on tests are mixed-positive: strong at building apps and games from scratch and fast on some coding tasks, but with slower generation (~33 tok/s anecdotes), average front-end aesthetics, and inconsistent results across accounts and sessions.

How do I call the API?

Subscribe to Token Plan (Singapore region), use base URL https://token-plan.maas.qwencloudapi.com/compatible-mode/v1 and model qwen4-max-preview. It works with Claude Code, Qwen Code, Cursor, Cline, Codex and other OpenAI/Anthropic-compatible tools.

Where can I check the latest status?

The Alibaba Cloud Model Studio docs, the QwenCloud Token Plan page, and the sources list below.

11Sources

Every fact on this page comes from the public sources below (verified Nov 12–13, 2026).

Official

English coverage & reviews

Chinese coverage & reviews