Alibaba Opens Qwen4-Max Preview With Million-Token Context and Image and Video Support

A larger workspace for Alibaba's next flagship AI

Alibaba has opened a preview of Qwen4-Max, the flagship model in its next generation Qwen family, with support for image and video inputs and a context window of up to one million tokens. Developers can access the model through Alibaba's Token Plan subscription, while early access also extends to the Qwen app, Qoder, QwenWork and QwenStudio.

Alibaba Cloud's Token Plan page lists the model as Qwen4-Max-Preview and identifies it as newly launched. Its arrival gives subscribers an early route into Qwen4 while the broader model family is still being introduced.

The release brings together three changes: room for much more material in a single request, the ability to receive visual information, and a stronger focus on programming and software tasks. Alibaba is presenting Qwen4 as a model intended to work through complicated jobs, including tasks that require several steps and the use of tools.

A limited promotion reduces Qwen4-Max preview Credits consumption to as little as 20% of the standard rate, equivalent to five times the usage, with a further nightly discount bringing consumption to 8% of the standard rate between 22:00 and 08:00 the following day (UTC+8). That makes pricing part of the launch pitch alongside capability, particularly for developers considering how much work they can afford to send to an AI model.

The central development is access to an early version of the flagship, rather than the release of every planned Qwen4 model. Alibaba's product plans also point to other versions aimed at different deployment needs, including one intended to run locally.

What one million tokens means

Qwen4-Max's context window can hold up to one million tokens. Tokens are the units an AI system uses to process information. For text, a token can be a word, part of a word or punctuation, so one million tokens does not mean one million words.

A context window is the amount of material a model can consider together while working on a request. Think of it as the working space available for instructions, documents, conversation history and other inputs. A larger window allows more of that material to stay together instead of being split across separate requests.

For programming, this could mean supplying more files from a software project at once. For office work, it could mean asking about a longer collection of documents while keeping the instructions and relevant passages in the same request. Those are practical reasons why context capacity matters to the tasks Alibaba is targeting.

Capacity and cost are separate considerations. A request that includes a large amount of material generally requires more processing than a short question. Developers evaluating the preview will therefore need to consider how much context their own tasks require alongside the subscription terms.

Images and video join the input

Qwen4-Max also accepts image and video inputs. That makes the preview multimodal, a term for AI that can receive more than one kind of information, rather than relying only on typed text.

Visual input fits several of the uses Alibaba has highlighted. A developer working on a website may need to show the model an interface as well as describe the code. A software task may involve information displayed on a screen. Receiving visual material gives the model another way to take in the task.

Video adds information that changes over time, while an image captures a single view. That difference can matter when the relevant detail is a sequence of actions rather than one screen. The preview's stated support concerns image and video inputs; Alibaba's broader promotional page separately lists Wan3.0-Video among its other models.

Programming and software tasks take priority

Alibaba is highlighting complex programming, frontend design, professional office work and the operation of professional software as priorities for Qwen4. The company's promotional descriptions say the model has delivered promising results in practical tasks across those areas.

Frontend design means building the parts of a website or application that people see and use, including pages, menus and controls. It combines visual decisions with programming, making it a useful example of why Alibaba is promoting coding and visual understanding together.

Agent capabilities are another focus. In this setting, an agent is an AI system that can use tools and work through a sequence of actions toward a goal. Instead of only returning an answer, it may need to choose a tool, examine the result and decide what to do next.

Alibaba's stated ambition is for Qwen4 to use tools autonomously and complete tasks with multiple steps. For a software workflow, the value would lie in connecting individual actions into a useful result. Users assessing that capability will need to pay attention to both the final output and the actions taken along the way, especially when a tool can change files or other working material.

Where users can try the preview

Token Plan is the subscription route identified for developer access. Alibaba Cloud markets the plan as access to Qwen models, models from other providers, tools and support within one subscription product.

The official page promotes Qwen4-Max-Preview alongside Qwen3.8-Max, Qwen3.8-Flash, Wan3.0-Video, DeepSeek-V4.1-Flash and GLM-5.2. That places the new flagship inside an existing collection of models rather than presenting it as a subscription with access to Qwen4 alone.

Early availability in the Qwen app, Qoder, QwenWork and QwenStudio broadens the rollout beyond the developer subscription. Someone already using one of those products can encounter the new model within an existing application, while developers have the Token Plan route for their own work.

The practical choice depends on the task. Trying the model inside an application offers a direct way to explore its responses. Evaluating it for a software workflow also requires checking how it handles that workflow's instructions, files and tools.

Who qualifies for the promotional rates

Alibaba Cloud's Token Plan documentation sets the preview window from November 11, 2026 at 10:00 through November 25, 2026 at 24:00 (UTC+8), with the model to be taken offline afterward. The preview is to remain unchanged during that window. The eligible model IDs are qwen4-max-preview and its alias qwen4.0-max-preview; both can be called, but bills display qwen4.0-max-preview. The documentation lists the model's capabilities as reasoning, vision understanding and text generation.

The Token Plan promotion describes the launch offer as an 80% saving on first-time Qwen4-Max preview calls. Its Double 11 campaign also advertises subscription prices starting at 50% off, a 78% increase in Credits and coverage of more than 20 flagship models. The page separately lists Qwen3.8-Max, DeepSeek and GLM among 17 models available through supported AI work tools.

For accounts that have never purchased Token Plan personal edition, Lite costs 19.5 yuan per month and Essential costs 39.5 yuan per month under the new-user offer, valid through November 30, 2026 at 23:59:59 (UTC+8). The personal plan is currently available only in the China North 2 (Beijing) region.

Calls must use the Token Plan-specific API key and Base URL available from the user's subscription page to draw down subscription Credits. A general API key will not qualify for that deduction. Personal-plan access is limited to interactive use in programming and agent tools such as Qoder, Claude Code, Cursor, Qwen Code and OpenClaw. Automated scripts, custom application backends and non-interactive batch calls are prohibited.

How Qwen4 reached this preview

Alibaba had already described Qwen4 as a model family based on a new architecture. Architecture means the underlying design of a model: how its components are arranged to process information. The company previously announced in September that training was under way.

Qwen3.8-Flash-Next had also offered an experimental version of the new architecture before the Qwen4-Max preview. The sequence shows that Alibaba introduced an experimental route to the design before opening early access to the next family's flagship.

• September: Alibaba announced that Qwen4, based on a new architecture, was in training.
• Before the flagship preview: Qwen3.8-Flash-Next offered an experimental version of that architecture.
• At the preview launch: Qwen4-Max became accessible through Token Plan and was integrated into the Qwen app, Qoder, QwenWork and QwenStudio.

The previously shown product plan also includes Qwen4 Flash, Plus and a 27B version aimed at local deployment. The 27B label refers to a model with approximately 27 billion parameters, the numerical values learned during training. Local deployment means running a model on an organisation's or user's own computing equipment.

A preview with a broader rollout still to come

Qwen4-Max gives developers and application users an early opportunity to assess the flagship on their own tasks. Its combination of a large context window, visual inputs and tool use ambitions is particularly relevant to work that involves several kinds of information or several stages of execution.

The launch details leave the date of full Qwen4 availability and release dates for the other planned versions unspecified. They also provide no detailed benchmark scores for the highlighted task categories.

For now, the concrete change is availability: users can begin trying Qwen4-Max through named Alibaba products. The broader plan extends beyond this flagship preview, with Flash, Plus and a model for local deployment still part of the proposed family.

Key Points

• Alibaba has opened Qwen4-Max-Preview through Token Plan and several Qwen applications.
• The flagship preview supports image and video inputs and up to one million tokens of context.
• Alibaba is targeting complex programming, frontend design, office work and software tasks involving tools.
• The preview runs from November 11 to November 25, 2026 (UTC+8), with promotional Credits consumption as low as 20% of standard and 8% overnight for eligible Token Plan calls.
• The planned Qwen4 family also includes Flash, Plus and a 27B version intended for local deployment.