Alibaba's Qwen team announced Qwen4 with a single social post and shipped a preview the same morning. No keynote, no technical report, no model card. The model ID is qwen4-max-preview, the flagship tier of the Qwen4 family.

What is actually specified

A 1M-token context window, unchanged from Qwen3.8-Max. Maximum output of 393,216 tokens, roughly three times the prior generation's 128K. Multimodal input: text, images up to 16 megapixels, video up to two hours and 2GB, with batches of up to 2,048 images or 64 videos. Hybrid thinking on by default, function calling, built-in tools and structured output.

The stated focus areas are deeper knowledge, better world perception, and stronger coding and agentic behavior, with four named scenarios: complex programming, professional office work, front-end design, and operating professional software. That last one is the interesting entry. It means the model is expected to click through interfaces and complete workflows, not just emit code.

What is not specified matters more

No parameter count. No published benchmarks. No general-availability date. Historically this family keeps Max closed behind the API and opens weights for the mid-tier later; third-party trackers expect Apache 2.0 weights for smaller sizes by December 20, but that is reporting, not an official commitment.

Without reproducible evaluations, buyers can judge whether the model works, not how it compares. That is a deliberate trade: Alibaba is asking users to evaluate on their own tasks instead of on a leaderboard.

Access and pricing

There is no standalone pay-as-you-go API yet. The only programmatic route is the Token Plan subscription, with first-party surfaces including Qwen Office, Qoder, the Qwen app and QwenStudio. Personal plans start at 39 yuan per month, with new-customer discounts down to 19.5 yuan, and preview-period discounts as steep as 90% off. The preview runs until November 25, after which the model identifier is retired.

Subscription rather than metered billing is a product decision. It lowers the cost of experimentation and keeps usage inside Alibaba's toolchain, at the price of making external cost comparisons impossible.

How to spend the two weeks

Bring a task that has never worked: turning a messy project directory into a runnable scaffold, or converting a two-hour meeting recording into a structured brief. Those expose boundaries quickly.

Test three things specifically. Whether long outputs drift halfway through a 300K-token document. How many consecutive steps it can complete in a professional application without supervision. Whether video understanding reads on-screen text accurately.

Do not migrate production. Interfaces, rate limits and pricing can all change inside a preview window. Run it in parallel and compare against your baseline.

If self-hosting is the goal, wait for open weights. The pattern in this family is a closed flagship with an open mid-size model following.

The broader signal is about release strategy, not capability. Withholding the spec sheet moves the argument from "what did you score" to "did it finish my job," which is where most enterprise buyers actually live.