A preview with an expiry date

Alibaba opened a preview of Qwen4-Max on Wednesday, the first model in its Qwen4 generation to reach users. The launch post from the Qwen team, published at 03:01 UTC on 11 November, said only that “Qwen4 is coming” and that the model was available on Alibaba’s Token Plan, plus early access through QwenWork, Qoder, the Qwen app and QwenStudio.

The model’s own documentation in Alibaba Cloud Model Studio is more specific, and more restrictive. The preview runs from 10:00 on 11 November to 24:00 on 25 November, Beijing time — fourteen days. After that, the documentation says, the model is retired. Pay-as-you-go access is not supported at any point.

An analogue wall clock on a plain wall
The preview is served from Alibaba Cloud rather than as downloadable weights. Illustrative image. Julia Filirovska · pexels · Pexels License

What the documentation does say

Qwen4-Max-Preview takes a context window of 1 million tokens and accepts text, images and video as input. Its maximum output is 393,216 tokens, and the documentation notes that this budget is shared between the model’s internal reasoning and the answer the user sees — a limit worth reading twice if you plan to use long chains of thought on long inputs.

What the documentation does not contain is as notable. There is no parameter count, no benchmark table, no model card and no link to weights. Alibaba has published nothing that would let anyone compare this model against Gemini 4 Carbon, GPT-6 or Kimi K3.1 Pro except by running it. That is a deliberate shape for a preview: the company is collecting evaluations from users rather than publishing its own.

The subscription is the gate

Reaching the model means buying a Token Plan subscription. Alibaba’s pricing page lists four individual tiers, currently discounted: Lite at $3 a month instead of $8, for 20,000 credits; Essential at $5 instead of $16, for 45,000; Standard at $18 instead of $25, for 80,000; and Pro at $43 instead of $51, for 200,000. The page tags qwen4-max-preview as “Limited-Time 80% Off” without stating its credit cost per token.

The same subscription also serves models from Alibaba’s competitors — DeepSeek V4.1 Flash and V4 Pro, and Zhipu’s GLM-5.2 and GLM-5.3 — and works with third-party tools including Claude Code, Cursor, Codex and Cline. Alibaba is selling a tap, not only a model.

Cooling fans mounted in a server rack
Alibaba has published no benchmark scores for the model. Illustrative image. panumas nikhomkhai · pexels · Pexels License

What to watch

Alibaba first showed the Qwen4 line — Max, Flash, Plus and an open 27-billion-parameter model — at its Apsara conference in September. Max is the first to ship in any form, and the other three have no announced dates.

The concrete marker is 25 November. Either a full Qwen4-Max replaces the preview with pay-as-you-go pricing and published numbers, or the model simply disappears and the generation’s first public outing leaves nothing behind but user reports.