Alibaba Releases Qwen4-Max-Preview: No Numbers, but a Test Window Until November 25

Alibaba has rolled out a preview of its new flagship Qwen4-Max, and did it in a "faster, higher, stronger" style: the announcement on X contains not a single number. No benchmarks, no comparison with the previous generation. What it does offer is a practical opportunity: while the model is in preview, you can cheaply try it out on your own tasks. The window is short, until November 25.

What we know about the model

Qwen4-Max-Preview was released on November 11, 2026. The documentation calls it a next-generation flagship multimodal model. The list of capabilities is familiar: function calling, internet search, built-in tools and responses strictly in JSON format. For those building agents and data-extraction pipelines, this is a basic set, nothing exotic.

What matters more is something else: the model is compatible with the OpenAI API and with Anthropic Messages. In practice, this means it can be connected to Claude Code, Cursor and other coding agents via a separate address and key. You will not have to rewrite your integration; it is enough to add one more endpoint in the settings.

What we don't know is also considerable. Code quality, speed, price per token: none of this appears in the materials we found. Per-token billing was not introduced at all for the preview, so it is impossible to calculate cost in the usual units.

How much it costs to try

Access is available through QwenCloud, Qwen APP, Qoder, QwenWork, QwenStudio and other Qwen platforms. The main route is the Personal Lite subscription at $6 per month. For new users, a price of $3 applies until November 30, but only if you have not previously purchased a personal plan.

The subscription includes 20,000 credits per month. Consumption depends on the model, the token volume, the reasoning mode and tool calls. This is not unlimited: an agent with long reasoning chains and dozens of function calls can burn through the package noticeably faster than a chat with a couple of questions.

During the preview, credits for this model are spent several times more slowly than usual, with a discount of up to 80%. At night Beijing time, from 22:00 to 08:00, it rises to 92%. In Moscow time, that is roughly 17:00 to 03:00. If you plan to run large batches of tests, it makes sense to schedule them in this window.

What to do with it

The main takeaway is simple: Qwen4-Max-Preview is so far a newcomer with no user track record. Qwen3.8-Max, which became generally available on August 3, remains a proven, stable model. Don't take claims without numbers on faith; compare for yourself.

A minimal plan until November 25:

• Collect 20–30 real tasks from your product: bugs from your repository, typical user requests, complex JSON schemas.
• Run them through both models with identical prompts.
• Compare not "liked it or not," but the share of valid JSON, the correctness of function calls, the number of agent steps and credit consumption.
• Separately test the model in Claude Code or Cursor: weak spots show up there fastest.

The limitations are obvious. The preview will be shut down on November 25, so it should not be put into production. Alibaba has not named a date for the full release, and it is unclear whether the price and terms will be preserved then. Without token billing, you cannot calculate the economics of your product on this model.

There is also context around the lineup. At the Apsara conference on September 22, Alibaba showed the Qwen4 family: Max, Plus, Flash and 27B. Timing and specifications were not given at the time. Whether the 27-billion-parameter version will be open for local deployment has not been officially confirmed, so it is too early to make plans for local deployment.

Bottom line: a release without numbers is a reason not to admire but to measure. If you already have an evaluation set for models, a couple of evenings and $3–6 are worth it. If you don't, this is a good reason to finally build one: it will be useful both for Qwen4 and for any next model.