Qwen4-Max Preview: Token Plan Only, Default Is xhigh, and It Retires on Nov 25

Attributed compile + Credit-ratio arithmetic on my box

Primary announcement: Qwen (@Alibaba_Qwen), posted 2026-11-11 02:01 UTC (2026-11-10, 21:01 America/Toronto). The X API returned the opening lines ("Qwen4 is coming… Qwen4-Max is now in preview on Alibaba’s Token") plus a media link; I did not reconstruct a longer caption.

Primary docs: qwen4-max-preview (Alibaba Cloud Model Studio) (last updated Nov 11, 2026).

Token Plan: Individual overview (QwenCloud), Individual overview (Alibaba Cloud), Individual quick start.

Call shape: Thinking, Text generation and visual understanding.

Discussion (not a source of numbers): HN: Qwen4-Max is now in Preview (45 points / 3 comments at check); HN: Qwen 4 Max Preview (14 points / 4 comments).

Local copies: /workspace/reputation/runs/39-qwen4-max-preview/.

Specs, prices, and promo rates below are vendor docs unless marked otherwise. This box did not hold a Token Plan or QwenCloud key and did not send a chat completion. The only local run is ratio arithmetic in credits.py. Artificial Analysis had no qwen4 listing in a models-index scrape, and likely slugs (qwen4-max-preview, qwen-4-max, qwen4-max) returned HTTP 404.

Alibaba's Qwen team opened a preview of Qwen4-Max today. The callable id is qwen4-max-preview. Official docs are unusually blunt about the terms: it is Token Plan only (no pay-as-you-go), the window runs from 10:00 on November 11 to 24:00 on November 25, 2026 (UTC+8), no changes will be made during that period, and the model will be retired after the preview ends. In America/Toronto that is November 10, 21:00 EST through November 25, 11:00 EST — about 14 days from this morning's check.

This is not a quiet docs typo. The same retirement language is on the model page and on both Token Plan Individual overviews. If you pin qwen4-max-preview in a tool config and walk away, that id is scheduled to go away on the 25th. There is also no public per-token price and no independent index score to shop against: thinking is on and xhigh by default, and thinking tokens share the 393,216 output budget with the visible answer. Those are the day-one facts, not a bench table that does not exist yet.

What shipped

Vendor docs, compressed:

QwenCloud's model guide also lists, for this id: up to 16M pixels per image, video up to 2 hours / 2 GB, up to 2,048 images by URL, and up to 64 videos. The Base64 image cap is marked "not specified upstream." Treat those as catalog limits, not something I measured.

The official announcement post was already doing the rounds this morning: at a 09:04 America/Toronto check it had 10,329 likes, 1,335 reposts, 775 replies, 930 quotes, 1,294 bookmarks, and 878,251 impressions. HN was much quieter (45 and 14 points on the two threads above). I am using those counts only as a snapshot of attention, not as a quality score.

Token Plan only — and the id is temporary

You cannot point a regular QwenCloud pay-as-you-go key at this model. Token Plan Individual quick start is specific about the pairing:

• Create or use a Token Plan key. Those start with sk-sp-. Regular keys start with sk-ws- and do not work with Token Plan.
• Use the Token Plan Base URL from the same console. QwenCloud's quick start lists the OpenAI-compatible endpoint as https://token-plan.maas.qwencloudapi.com/compatible-mode/v1 and the Anthropic-compatible endpoint as https://token-plan.maas.qwencloudapi.com/apps/anthropic. The Alibaba Cloud model page lists a dedicated Base URL of https://token-plan.ap-southeast-1.maas.aliyuncs.com and tells you to take the complete per-protocol path from the Token Plan guide. Copy the pair your console shows; do not mix a Token Plan key with a region-coded pay-as-you-go URL.
• Set the model to qwen4-max-preview. The alias qwen4.0-max-preview is also accepted, but invoices show the alias.

Token Plan itself is a Credits subscription, not a meter you can turn on for one call. The Individual table on the fetched overview:

Two access constraints are easy to miss and are not optional fine print. First, Token Plan Individual is sold for Singapore, with Global inference, and the docs say prompts and outputs involve cross-border data transfer. The purchase entry is hidden if the console is set to another region. Second, the Individual terms of use limit the key to interactive use inside programming and agent tools (the overview names Claude Code, Cursor, Qwen Code, Qoder, and OpenClaw). They explicitly disallow automation scripts, custom application backends, and non-interactive batch calls. A first look at Qwen4-Max is a Cursor or Claude Code session, not a new production backend on a preview id that retires in two weeks.

A Chat Completions call should pin the model, set effort yourself, and leave room in the output cap for thinking. The shape follows QwenCloud's Token Plan Base URL and Thinking examples; this box did not execute it:

curl -X POST "https://token-plan.maas.qwencloudapi.com/compatible-mode/v1/chat/completions" \ -H "Authorization: Bearer $TOKEN_PLAN_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "model": "qwen4-max-preview", "messages": [ {"role": "user", "content": "Outline a safe rollback plan for a failed FastAPI migration, in six bullets."} ], "max_tokens": 8192, "enable_thinking": true, "reasoning_effort": "medium" }' 

enable_thinking and reasoning_effort go in the JSON body on a raw request. If you use the official Python SDK, they are not standard OpenAI fields — pass them in extra_body. On the Responses API the depth knob is reasoning.effort. thinking_budget is supported on Chat Completions and DashScope, not Responses, and must not be combined with reasoning_effort on this model.

Default is xhigh, and thinking shares the 393K cap

Three defaults will surprise you if you treat this like an older Qwen id.

Thinking is on, at xhigh, unless you change it. The model is hybrid, so you can turn thinking off with enable_thinking (or the Responses equivalent). If you set neither reasoning_effort nor thinking_budget, docs say the default applies — and the model page's default is xhigh. The published ladder is only low, medium, and xhigh. There is no minimal / high / max label here; do not copy a Muse or OpenAI effort string and assume it is valid.

The output budget includes thinking. Starting with qwen4-max-preview, max_tokens is equivalent to max_completion_tokens. Both cap thinking plus the final answer. A limit you previously sized for the visible reply alone can truncate the answer because thinking has already spent part of the budget. Older Qwen models keep their previous behavior. Thinking tokens are also billed as output tokens. Log the reasoning usage your endpoint returns; this box did not measure a split.

Multi-turn Chat Completions enables preserve_thinking by default. Send historical thinking back in each assistant message's reasoning_content field. Do not concatenate it into content. If you drop that field, you are not doing what the default path expects.

The 20% / 8% Credit promo — and why Toronto daytime is the cheap window

There is still no public Credits-per-token table for qwen4-max-preview. Docs say actual consumption is dynamic (model, tokens, thinking mode, tool calls) and that you should read Model usage on the Token Plan page. What they did publish is a preview promotion: Credits consumption is as low as 20% of the standard rate (5× usage); calls between 22:00 and 08:00 UTC+8 get an additional 60% off, i.e. 8% of the standard rate.

That "night" window is the interesting part if you are in America/Toronto. 22:00–08:00 UTC+8 is 09:00–19:00 EST — ordinary weekday working hours here. The cheaper 8% band is Beijing night and Toronto day, not the other way around.

I ran credits.py on those official percentages and the published quotas (arithmetic, not an invoice and not a unit price):

day rate vs standard: 20% -> 5.0x quota equivalent night rate vs standard: 8% -> 12.5x quota equivalent night vs day promo: 0.40x (40% of the already-discounted day rate) plan list USD quota day-equiv night-equiv Lite $ 6/period 20,000 100,000 250,000 Essential $ 10/period 45,000 225,000 562,500 Standard $ 18/period 80,000 400,000 1,000,000 Pro $ 43/period 200,000 1,000,000 2,500,000 Credit Pack $ 15/period 20,000 100,000 250,000 

A Lite quota of 20,000 Credits behaves like 100,000 standard-rate Credits in the 20% band and like 250,000 in the 8% band — if your calls actually bill at those promoted rates. The 8% window is 40% of the already-discounted daytime promo, not 8% on top of a mystery unit price I do not have. After November 25, the preview id is scheduled to retire, so do not plan a month of production traffic against this multiplier.

What this box actually did

preflight 2026-11-11 ~09:02–09:04 America/Toronto NO TOKEN PLAN KEY / NO QWENCLOUD KEY / NO CHAT COMPLETION HTTP 200: Alibaba model page; QwenCloud Token Plan Individual overview and quick start; Alibaba Token Plan Individual overview; Thinking; text-generation models; pricing hub; OpenAI chat reference HTTP 404: AA /models/qwen4-max-preview, /models/qwen-4-max, /models/qwen4-max AA models index scrape: no "qwen4" string X API: official announcement post + public_metrics (counts above) HN Firebase: items 50449130 (45/3) and 50451192 (14/4) Parsed: Token Plan only; retires after 24:00 Nov 25, 2026 UTC+8 (Nov 25, 11:00 America/Toronto); default thinking xhigh; shared 393,216 output cap; max_tokens == max_completion_tokens; promo 20% / 8% of standard Credits; sk-sp- key + Token Plan Base URL Local run: credits.py ratio arithmetic only NOT tested: authenticated completion; Credit burn per token; tool calls or video/image input; Cursor / Claude Code / Qwen Code; whether the two documented Base URLs are interchangeable 

Qwen4-Max Preview is a real, short, subscription-gated look at the next Qwen flagship: 1M context, multimodal, familiar API shapes, and a Credit promo that is unusually cheap during Toronto business hours. For a Toronto product team on 2026-11-11, the rational first move is subscribe to Token Plan if you don't have it, pin qwen4-max-preview in an interactive coding tool, set reasoning_effort below xhigh for routine work, raise max_tokens so thinking does not eat the answer, and take notes before November 25 — not to treat a retiring preview id as the new production default.

YongBo Yu — Toronto AI engineer (agent systems / LLM workflows). Site: yongbo-yu.vercel.app · GitHub: YongBoYu1