On November 11, 2026, Alibaba's Qwen team announced the launch of Qwen4, its next-generation AI model family, and opened a limited preview of its flagship model, Qwen4-Max. The model supports a 1-million-token context window and is presented as having enhanced capabilities for coding and AI agents. It is available through Alibaba Cloud's flat-rate API service Token Plan, as well as QwenWork (a development support environment), Qoder (a coding tool), Qwen APP (the consumer app), and QwenStudio (an integrated development environment).
However, this release is not the start of general availability. According to the official documentation, the Token Plan preview ends on November 25, and the plan does not allow the model to be built into automation scripts or the backends of in-house applications. It lets developers try the new model's performance early, but the format is not suited to full-scale adoption in existing systems. Below, we go through Qwen4-Max's technical specifications and terms of use, and look at what developers can actually evaluate in this preview.
Qwen4-Max Debuts Early on Five Services
Qwen's official X account announced the launch of the Qwen4 model family at 2:01 a.m. UTC on November 11, 2026 (11:01 a.m. the same day in Japan). The full series has not been officially released, however; what is available now is a preview of Qwen4-Max. The announced access points are five services: Token Plan, QwenWork, Qoder, Qwen APP, and QwenStudio.
Each service differs in its target users and terms of use. QwenStudio offers a free plan, but features and usage allowances are not the same across all services. Pricing structures and terms differ depending on whether you try the model through a browser or dedicated client, or integrate it into development tools via the API, so each needs to be checked individually.
The Qwen team says the new model has improved performance in complex coding, work requiring specialized knowledge, front-end development, and professional software development. It also says breadth of knowledge, world understanding, and agent capabilities have improved. At this point, however, these are the developer's own claims, and no independent third-party comparative benchmarks have been published.
The rollout appears designed to let a wide range of users, from general consumers to software developers and corporate development teams, try the new model. API access, however, requires a Token Plan subscription and is restricted in both duration and permitted uses. In its current form, Qwen4-Max should be seen as a preview for evaluating features and performance, not a finished product intended for production systems.
1M-Token Context and Up to 393,216 Output Tokens
The API model identifier for the Qwen4-Max preview is "qwen4-max-preview," and the official documentation also lists an alias, "qwen4.0-max-preview." It cannot be called from the standard pay-as-you-go API and is limited to Token Plan subscribers.
The most notable published specifications are a context window of 1 million tokens (1,000,000 tokens) and a maximum output limit of 393,216 tokens. The model can take in large source code bases and long documents, and can also generate long responses. However, the maximum output token count includes both the Thinking tokens used for the model's internal reasoning and the final Answer tokens. If a lot of tokens are spent on reasoning, fewer remain for the final answer.
Input supports text, images, and video, while output is text. The model also supports function calling, built-in tools, and structured output using tools such as JSON Schema. "Hybrid thinking," which switches between a reasoning mode and a direct-answer mode depending on the situation, is enabled by default.
For video input, the limits are up to 2 hours, a file size of 2 GB or less, and up to 64 videos per request. These are upper limits on the data that can be submitted, however, and do not guarantee that the model can accurately understand long videos. Likewise, for the 1-million-token context window, the length of input that can be accepted should be distinguished from the ability to accurately retrieve necessary information from a long context.
As of November 11, Qwen4-Max's total parameter count, objective benchmark results, and the timing of general availability (GA) had not been disclosed. On August 26, 2026, the Qwen team released "Qwen3.8-Flash-Next" on GitHub as an early preview of its next-generation architecture, presenting a design that combines Gated DeltaNet (GDN) and Qwen Sparse Attention. It is not clear, however, whether this architecture is used in Qwen4-Max.
These specifications suggest an emphasis on development support that handles large codebases and long documents, and on AI agents that work with external tools. Actual coding accuracy and reasoning reliability, however, cannot be judged from the published specifications alone. Understanding the performance gap with existing models requires testing in which they process the same real tasks under the same conditions.
| Evaluation access route | Service positioning | Pricing / allowance | Main terms of use | Expected evaluation use |
|---|---|---|---|---|
| Qwen APP / QwenWork / Qoder | Consumer and business chat environments, development tools | Subject to each service's terms (some free allowances) | Follows the rules of each app or client | Trying prompts; checking conversational performance and coding assistance |
| QwenStudio | Integrated AI workspace | Free plan available (per PC Watch) | Dedicated usage allowance applies; separate from API key issuance | Trying prompts and tool integration |
| Token Plan (individual) | Flat-rate API service | $6–$43 per month, 20,000–200,000 Credits | Limited to use in interactive development tools; embedding in in-house backends prohibited | Verifying code generation and editing with tools such as Cline and Cursor |
| Token Plan (team) | Flat-rate API service for organizations | $20–$180 per seat per month, 25,000–250,000 Credits | One user and a dedicated API key per seat; key sharing prohibited | Joint evaluation by multiple developers |
As shown, the terms of use differ greatly depending on how you try Qwen4-Max. Qwen APP and similar options are fine for simply checking general response quality, but evaluating integration with actual development tools requires Token Plan. When using the API in particular, check the preview end date and the terms of use in advance.
Token Plan Access Runs Only Until November 25; Mind the Pricing and Terms
The most important constraint of this preview is that availability on Token Plan is limited to about two weeks. The official documentation states that the availability period runs from 10:00 on November 11, 2026 to 24:00 on November 25 (UTC+8, China Standard Time). In Japan time, that is 11:00 on November 11 to 1:00 on November 26. It also states that the model will be "retired" after the period ends, so continued use on the plan is not guaranteed, at the very least.
The individual Token Plan is a monthly flat-rate subscription offered in four tiers: Lite, Essential, Standard, and Pro. The regular prices are $6 per month for Lite, $10 for Essential, $18 for Standard, and $43 for Pro. For new users, limited-time discounted prices of $3 for Lite and $5 for Essential are also offered.
Each plan is allotted a fixed amount of Credits each month, ranging from 20,000 Credits for Lite to 200,000 Credits for Pro. Credit consumption varies with the model used, the number of input and output tokens, the amount of processing in thinking mode, and the number of tool calls. Even with a fixed monthly fee, the number of tasks you can process is not fixed, so actual consumption should be checked during evaluation.
Another point to note is that the permitted uses of the individual Token Plan are strictly limited. The plan is intended for interactive use through IDEs, coding assistance tools, AI agents, and the like. The official documentation prohibits bulk processing by automation scripts, use as the permanent backend of an in-house application, and embedding in non-interactive batch API processing. Being able to obtain an API key does not mean the model can be used like a typical commercial API.
In addition, Token Plan can be purchased and used only in the Singapore region. The API key and connection endpoint are also specific to Token Plan and are not compatible with the standard pay-as-you-go API. Even if a official release of Qwen4-Max is later offered through the pay-as-you-go API, the current connection settings may not carry over as they are.
Companies considering use should also pay attention to data handling. According to QwenCloud's official documentation, in the global environment, entered prompts and generated results may be transferred across borders as part of inference processing. Users are responsible for confirming compatibility with their own compliance requirements and applicable laws and regulations. When handling data that includes confidential or personal information in particular, processing regions and storage conditions must be checked.
Given these conditions, it is reasonable to treat the Token Plan preview as an opportunity to test fit with development workflows, not as a step toward moving to production. API and agent-oriented features are available, but continued availability after the preview is not guaranteed, and backend use is restricted. Until the terms for the official release are clear, it is best to avoid decisions to replace existing production systems with Qwen4-Max.
Four Things to Test While Waiting for the Official Release
For developers, these roughly two weeks are a chance to gauge what Qwen4-Max can really do. What matters is not building the model into existing systems but comparing it with the models currently in use and identifying which tasks might see improvement. It makes sense to use the interactive development tools permitted under Token Plan and give the model tasks close to everyday work.
The preview period restriction is the same for the team Token Plan. Team plans are priced at $20 per seat per month (25,000 Credits), $75 (100,000 Credits), and $180 (250,000 Credits). Each seat is assigned a dedicated API key, so multiple developers can evaluate the model in their own environments. Even with an organizational contract, however, the scheduled November 25 end of availability is not extended.
In actual testing, the following four areas are important.
- Understanding and modifying large codebases: Try grasping dependencies across multiple files, modifying existing code, and refactoring. The key evaluation point is whether changes remain consistent even when a long context is provided.
- Tool operation by AI agents: Give tasks that use external tools, such as file system operations and terminal command execution. Check whether function-call formats are correct and whether multiple operations are executed in the proper order.
- Processing long-form and multimodal input: Input large volumes of technical documents, multiple screenshots, long videos, and so on, and check whether the model can accurately extract the necessary information. In particular, it matters whether the model overlooks information in the middle of a long input.
- Generating structured data: Specify a format such as JSON Schema and verify whether the output matches the requested form. Check whether the format is respected even with complex hierarchies or strict type specifications.
In the evaluation, it is a good idea to record not just whether a task was completed, but also the failure rate, the number of human corrections or interventions, response time (latency), and Credit consumption. Comparing with existing models using the same inputs, the same tools, and the same evaluation criteria makes it easier to see not only differences in model performance but also differences in the time and cost that arise in real work. Note that large volumes of automated API calls may violate Token Plan's terms of use, so testing methods must also stay within the rules.
To decide on production deployment, you need to confirm whether an official model will be offered after the preview ends, how pay-as-you-go API pricing will be set, and whether embedding in commercial backends will be permitted. Data processing and storage regions and ongoing support arrangements are also essential conditions for enterprise operations.
Qwen4-Max stands out as a next-generation model with a 1-million-token context window, multimodal input, and agent-oriented features. At present, however, there are no independent evaluation results to back up its performance, and availability on Token Plan is time-limited. The realistic approach is to verify practical performance during the limited preview period, and to decide on adoption once the official release's specifications and commercial-use terms are clear.