Alibaba added qwen4-max-preview to Model Studio on November 11, 2026, and the only way to call it through an API is a Token Plan subscription. Model Studio release log The model page says pay-as-you-go access is not supported, that the preview runs from 10:00 on November 11 to 24:00 on November 25 (UTC+8), that no changes will be made during that period, and that the model will be retired when the preview ends. Qwen4-Max-Preview model page By our conversion the end is 11:00 am US Eastern on November 25. That is about two weeks of hands-on access through a coding-tool subscription, with no per-token price and no benchmark or weights that we could find. Nothing published yet supports planning a product around it.
S5 Labs has not used the model. Everything below comes from pages published by Alibaba and the Qwen team, read on November 12, and from our earlier coverage. Alibaba describes Qwen4-Max-Preview as the next-generation flagship multimodal model in the Qwen family, aimed at coding, office productivity, frontend design and complex agentic tasks.
Access needs a Token Plan key, and scripted use breaks the terms
Per the model page, you subscribe to Token Plan Personal or Team, get a dedicated API key, use the Base URL https://token-plan.ap-southeast-1.maas.aliyuncs.com and set the model ID to qwen4-max-preview. The Personal quick start lists the OpenAI-compatible and Anthropic-compatible endpoints and says Token Plan keys start with sk-sp- and are not interchangeable with standard Model Studio keys. The alias qwen4.0-max-preview works too and is the name bills show. Token Plan Personal overview
The Personal overview limits the plan to interactive use inside coding and agent tools such as Claude Code, Cursor and Qwen Code. Automation scripts, custom application backends and non-interactive batch calls are prohibited, and a violation can bring a suspended subscription or a banned key. The Team overview also limits its plan to interactive use and bars automated scripts and application backends, so a scripted evaluation harness is outside the terms of both editions.
The English docs offer Token Plan in Singapore only. The Personal overview adds that the deployment mode is Global, so prompts and outputs cross borders, with compliance left to the subscriber. The English Team overview says conversation data is not used for model training; the English Personal pages say nothing about data use in either direction. Alibaba’s Chinese model page lists the same preview with a Beijing host, under a separate Beijing-only plan priced in yuan, whose Personal overview says model inputs and outputs are used for service improvement and model optimization.
The Qwen account’s launch post also offers early access in QwenWork, Qoder, the Qwen app and QwenStudio; it is our only source for those four.
Plans list from $6 a month, and Credits do not convert to dollars
Monthly list prices in US dollars from the English Personal and Team pages on November 12:
| Plan | Price per month | Monthly Credits |
|---|---|---|
| Personal Lite | $6 ($3 new-user price) | 20,000 |
| Personal Essential | $10 ($5 new-user price) | 45,000 |
| Personal Standard | $18 | 80,000 |
| Personal Pro | $43 | 200,000 |
| Team Standard | $20 per seat | 25,000 per seat |
| Team Advanced | $75 per seat | 100,000 per seat |
| Team Premium | $180 per seat | 250,000 per seat |
The new-user price runs until 23:59:59 on November 30 (UTC+8) and applies only to accounts that have never bought a Personal subscription. Extra Credits cost $15 per 20,000 on Personal. The Personal Edition does not currently support unsubscription. Per the Personal FAQ, calls are blocked when Credits run out, no pay-as-you-go charges apply, and unused quota does not roll over.
The Personal overview says Alibaba raised all four Personal quotas and cut the Pro price on November 9. That page does not give the earlier figures; it points to a separate price reduction notice for details. The Token Plan prices in our July article on the Qwen3.8-Max preview are out of date.
Alibaba says the preview’s Credits consumption is as low as 20% of the standard rate, and 8% for calls between 22:00 and 08:00 (UTC+8), which by our conversion is 9:00 am to 7:00 pm US Eastern. It reserves the right to change promotions. Personal overview We found no published standard rate for this model. The Team overview says Credits per request depend on factors such as model type, token usage, thinking mode and tool calls. Nothing on these pages turns a plan price into a dollar-per-token figure for this model, so the console’s usage details are the only cost readout.
The pay-as-you-go price list has no Qwen4 entry in any region. qwen3.8-max is on it at $2 input and $6 output per million tokens (International), the rate we reported at its general availability.
The limits are documented, and the output cap now includes thinking
| Item | What Alibaba’s model page states |
|---|---|
| Context length | 1M tokens |
| Maximum output | 393,216 tokens, shared by thinking and the final answer |
| Input | Text, images and video; no audio input listed |
| Thinking | On by default; reasoning_effort takes low, medium or xhigh (the default) |
| Tools and formats | Function calling, web search, built-in tools; JSON Object and JSON Schema |
| APIs | OpenAI Chat Completions, Responses, Anthropic Messages |
| Rate limits (Personal FAQ) | No TPM or RPM published; limits adjust with load and cannot be increased; queuing possible at peak hours |
Starting with this model, max_tokens in Chat Completions is equivalent to max_completion_tokens and caps thinking plus the answer together, so a limit sized for the answer alone can truncate output. Alibaba’s compatibility log records the change under November 2. Client defaults sit far below the 393,216 ceiling: QwenCloud’s tool guides give 32,768 for Qwen Code and 32,000 for OpenCode.
The Chat Completions reference maps low, medium and xhigh to thinking_budget values of 4,096, 16,384 and 262,144. Setting reasoning_effort and thinking_budget together returns an error.
We found no benchmark, no weights and no general-availability date
We found no Alibaba benchmark for the model. The model page and release log contain no scores, and the launch post promises “Stronger coding and agentic capabilities” without data.
We found no Qwen4 weights, no license and no Alibaba statement about either, which is not proof that no statement exists. The Qwen organization on Hugging Face had no Qwen4 repository on November 12; its newest by creation date was Qwen-Image-2.1-Turbo, from October 9. Hugging Face API We checked only the Qwen organization; a Qwen4 name on a repository under another account does not by itself make it an Alibaba release.
The previous flagship has an open checkpoint. The Qwen3.8-2.4T-A95B card says hosted Qwen3.8-Max is based on it, and its license requires model-as-a-service and AI work assistant businesses with more than US$50 million in aggregate revenue over any 12 months to obtain a separate license from Qwen, as we covered in August. We found no Alibaba statement that Qwen4-Max will get an open checkpoint. Our Open Models guide keeps Qwen 4 under announced, not released. The hosted preview replaces its earlier in-training status, but Qwen 4 still fails the guide’s check for a self-hosting plan: weight files, a model card and a license.
The Alibaba pages we read name no other Qwen4 model, no parameter count and no general-availability date. The company’s September 22 press release said Qwen 4 was in training and that the upcoming Qwen 4.5 and Qwen 5 series were projected to scale up to 5 to 10 trillion parameters.
The Qwen3.8-Flash-Next card calls that open model an experimental preview of the architecture that will underpin Qwen4, which our August article and technical comparison cover. The Qwen4-Max-Preview page says nothing about architecture.
Alibaba says only that the model will be retired after the preview. It names no successor and does not say whether the ID will be rerouted or the promotion continued.
What to test before November 25, and what to wait for
In July we said the Qwen3.8-Max preview was not worth evaluating. This one has documented limits, a dated end and an entry price of $6 or less, so a short hands-on look is cheap.
Worth testing by hand, inside a supported tool:
- A fixed handful of your own hard coding or agent tasks, run on your current model and on
qwen4-max-preview. - Credits consumed per session at
xhigh,mediumandlow, read from the console. At a promotional rate these compare settings with each other and say nothing about a later price. - Long outputs against your client’s cap and the shared thinking-plus-answer budget.
- Image and video input, if your workflow uses them.
Before planning around Qwen 4, wait for a pay-as-you-go price, a non-preview model ID, published rate limits, an independent benchmark, and a statement on weights and license. Until then, keep the model ID out of anything that has to work after the preview, quote no per-token cost, and buy the smallest plan that covers your test. Personal plans cannot currently be unsubscribed, and Alibaba says the model will be retired after the preview ends on the 25th.