Alibaba opens Qwen4-Max preview on Token Plan with 1M context until Nov 25
Updated: 3 minutes ago
Alibaba has opened a limited preview of Qwen4-Max, the first model of its Qwen4 generation, and it has done so in an unusual way: the model is available only through Token Plan, the flat-rate monthly subscription that Alibaba Cloud Model Studio sells for use inside coding and agent tools. The model ID is qwen4-max-preview, with the alias qwen4.0-max-preview, and Alibaba Cloud's documentation page for it was published on November 11, 2026.
The preview runs from 10:00 on November 11 to 24:00 on November 25, 2026, China Standard Time (UTC+8), which corresponds to 02:00 UTC on November 11 through 16:00 UTC on November 25, or 9:00 p.m. Eastern on November 10 through 11:00 a.m. Eastern on November 25. Alibaba states that the model will not change during that window and will be retired once the preview ends, so developers have roughly 14 and a half days to test it before it disappears from the catalog.
On paper the model sits at the top of Alibaba's lineup. The documentation lists a 1M-token context window, a maximum output of 393,216 tokens shared between reasoning and the final answer, text, image and video input, and compatibility with three API formats: OpenAI Chat Completions, the OpenAI Responses API and the Anthropic Messages API. Pay-as-you-go access is explicitly not supported, and Alibaba has not published a per-token price, a parameter count, benchmark results or a general availability date for the model.
··········
QWEN4-MAX-PREVIEW SPECIFICATIONS AND API FEATURES.
The Model Studio page describes Qwen4-Max-Preview as a reasoning model with thinking enabled by default. Developers can set the depth of reasoning with the reasoning_effort parameter, which accepts low, medium and xhigh, with xhigh as the default. Alibaba warns that reasoning_effort and thinking_budget should not be set in the same request, a constraint for anyone porting code from earlier Qwen models that relied on an explicit thinking budget.
The feature list covers deep reasoning, function calling, web search and other built-in tools, and structured outputs in both JSON Object and JSON Schema modes. Input can be text, images or video, and the output is text. The page gives no separate figure for maximum input, so the practical ceiling for a prompt is the 1M-token context minus whatever space the request reserves for reasoning and the answer.
Two implementation notes in the documentation affect existing integrations. In Chat Completions, max_tokens behaves the same as max_completion_tokens for this model, and because thinking consumes part of that budget, a low limit can cut the visible answer short. The same endpoint also enables preserve_thinking by default, which means that in multi-turn conversations the model's earlier reasoning has to be returned in the reasoning_content field of each assistant message, never concatenated into the content field. Clients that strip or merge reasoning text, a common shortcut in simple chat wrappers, will need adjusting.
Access goes through a dedicated Token Plan endpoint in Singapore, with the base URL https://token-plan.ap-southeast-1.maas.aliyuncs.com, and API keys issued for Token Plan are separate from standard Model Studio keys. A developer who already pays for Qwen through the regular pay-as-you-go API cannot point that key at the new model.
··········
QWEN4-MAX-PREVIEW VS QWEN3.8-MAX: CONTEXT, OUTPUT AND ACCESS.
The closest point of comparison is Qwen3.8-Max, the flagship that Alibaba released in September with the snapshot qwen3.8-max-0902 and covered in our Qwen3.8-Max explainer. Both models share a 1M-token context window and accept text, image and video input. The larger difference lies in output length and in how they are sold.
........
Item | Qwen4-Max-Preview | Qwen3.8-Max |
Model ID | qwen4-max-preview | qwen3.8-max (snapshot qwen3.8-max-0902) |
Context window | 1M tokens | 1M tokens |
Max output | 393,216 tokens incl. thinking | 131,072 tokens |
Input | Text, image, video | Text, image, video |
Reasoning control | reasoning_effort: low, medium, xhigh | Thinking mode, no effort levels listed |
Pay-as-you-go API | Not supported | $2.00 input / $6.00 output per 1M (International) |
Token Plan | Yes, preview only | Yes |
Availability end | Retired after Nov 25, 2026 | No end date published |
........
The output ceiling of Qwen4-Max-Preview is three times that of Qwen3.8-Max, although the newer model counts reasoning tokens inside that budget, so the gain in visible answer length depends on how much the model thinks. For long agent runs, large code generation tasks and multi-file edits, a higher ceiling reduces the number of continuation calls a tool has to make.
Pricing cannot be compared directly. Qwen3.8-Max has a published pay-as-you-go rate on the International deployment of $2.00 per 1M input tokens and $6.00 per 1M output tokens, with implicit cache reads at $0.25 per 1M, according to the Model Studio model page checked on November 11, 2026. Qwen4-Max-Preview has no token price at all, because the only way to call it is to spend Token Plan credits, and Alibaba has not disclosed the standard credit rate per token for the new model.
··········
[[ADPLUS_300x250_1]]
QWEN4-MAX-PREVIEW ACCESS, TOKEN PLAN PRICES AND PREVIEW DISCOUNTS.
Token Plan is sold as a Personal Edition and a Team Edition, both offered in the Singapore region on Alibaba Cloud's international site and both updated on November 11 to list the new model. Alibaba's Chinese site sells a separate yuan-priced Personal Edition in the China (Beijing) region that also lists Qwen4-Max-Preview. The international prices are shown in the table below, with the price per 1,000 credits computed from Alibaba's published figures.
........
Plan | Monthly price | Credits per month | Cost per 1,000 credits |
Personal Lite | $6 ($3 for new users) | 20,000 | $0.30 ($0.15 new users) |
Personal Essential | $10 ($5 for new users) | 45,000 | $0.222 ($0.111 new users) |
Personal Standard | $18 | 80,000 | $0.225 |
Personal Pro | $43 | 200,000 | $0.215 |
Team Standard (per seat) | $20 | 25,000 | $0.80 |
Team Advanced (per seat) | $75 | 100,000 | $0.75 |
Team Premium (per seat) | $180 | 250,000 | $0.72 |
........
The new-user prices apply until 23:59:59 on November 30, 2026 (UTC+8) and only to accounts that have never bought a Personal Edition subscription. On a per-credit basis, the Essential plan comes out slightly cheaper than Standard, and Pro is the cheapest Personal tier. Team seats cost more than three times as much per credit as the Personal plans, a premium for shared administration rather than for extra model access.
For the preview window Alibaba is discounting the model heavily. Credit consumption for Qwen4-Max-Preview is "as low as 20% of the standard rate," which Alibaba describes as five times the usual usage. Calls made between 22:00 and 08:00 China Standard Time receive a further 60% reduction, bringing consumption down to 8% of the standard rate. In practice, a Lite subscriber's 20,000 monthly credits stretch as far as 100,000 credits of standard-rate usage during the day and 250,000 at night, and a Pro subscriber's 200,000 credits cover the equivalent of 1 million and 2.5 million standard-rate credits respectively. Those multipliers come from Alibaba's own percentages; the absolute number of tokens a credit buys on this model remains undisclosed.
The discounted window has a useful quirk for readers in the Americas. The 22:00 to 08:00 period in China corresponds to 14:00 to 00:00 UTC, which is 9:00 a.m. to 7:00 p.m. Eastern Time while US daylight saving time is over. US developers testing during their normal working day are therefore inside Alibaba's night-rate window for most of it.
··········
WHAT QWEN4-MAX-PREVIEW MEANS FOR DEVELOPERS USING TOKEN PLAN.
Token Plan is designed for interactive use, and the rules apply to the new model as much as to the rest of the catalog. Alibaba's terms for both editions say the subscription may not be used for automation scripts, custom application backends or non-interactive batch calls, and violations can lead to suspension of the subscription or revocation of the API key. Personal plans may not be shared, each verified identity is limited to one subscription per edition, and the same key may be used on several devices owned by the subscriber. Teams that want to put Qwen4-Max behind a production service have no supported route to do so during the preview.
The supported clients are coding and agent tools. Alibaba lists Claude Code, Cursor, Qwen Code, Qoder and OpenClaw as compatible, and Qwen4-Max-Preview's support for the Anthropic Messages API and OpenAI's Chat Completions and Responses formats is what lets those tools talk to it after a change of base URL, API key and model name. Developers who already run Claude Code or Cursor against a Token Plan key for Qwen3.8-Max can switch the model ID to qwen4-max-preview without changing anything else.
Readers planning a test should keep three practical limits in mind. The model is retired at the end of the preview, so any workflow built around it will need a fallback such as qwen3.8-max, which remains on the plan. Thinking is on by default at the xhigh setting, which uses more of each request's output budget and more credits per call than the low or medium settings. And because the Singapore endpoint is the only international access point, latency from North America and Europe will reflect that routing. For background on how Qwen models handle long inputs and conversation memory, our guide to the Qwen context window and token limits covers the earlier generations, and our report on Apple Intelligence using Alibaba's Qwen in China explains where Qwen sits in Alibaba's consumer partnerships.
··········
QWEN4-MAX-PREVIEW: CONFIRMED FACTS AND OPEN QUESTIONS.
Alibaba's documentation confirms the essentials of the release: a model called qwen4-max-preview with a 1M-token context, a 393,216-token output ceiling that includes reasoning, three reasoning effort levels, multimodal input, compatibility with OpenAI and Anthropic API formats, and access restricted to Token Plan subscribers between November 11 and November 25, 2026, with credit consumption cut to 20% of the standard rate and to 8% during China's night hours. The subscription prices, from $6 a month for Personal Lite to $180 per seat for Team Premium, are published on Alibaba Cloud's own pages.
The documentation leaves the most important technical questions open. Alibaba has not published benchmark results, a model card, a parameter count, the standard credit rate per token for this model or a pay-as-you-go price, and it has not said whether the final Qwen4-Max will be released as open weights or what other Qwen4 sizes will follow. Developers who want to judge the model on their own code have a fixed two-week window and a subscription that costs as little as $3 for new users to do it, after which Alibaba will have to publish the full release terms for Qwen4.
··········
FOLLOW US FOR MORE.
·····
DATA STUDIOS
·····
datastudios.org
