A flagship with an expiry date
Most model previews are open-ended. This one is a window: November 11 at 10:00 to November 25 at 24:00, China time, and then, in TechNode's words, this preview version goes offline. Alibaba is giving developers a fixed period with the first Qwen4 model, at event pricing PC Watch reports as starting at $3 a month for new Token Plan users, with no commitment in the announcement about the model's identifier, limits or pricing after that date.
That structure changes what the preview can be used for. It is not a base to build a workflow on: nothing about it is guaranteed to survive November 25. It is a bounded evaluation — a chance to run the new architecture against your own tasks and keep your own record of the results, and nothing more durable than that record.
What is actually documented
Separate the launch into what Alibaba said, what trade press reported, and what is still missing.
Alibaba's own announcement is short. It says Qwen4 is coming, that the Max preview is live on the named products, and that knowledge, coding and agent capabilities have improved. The post contains no numbers at all: no parameters, no scores, no context length. PC Watch states the position plainly — as of November 11, Qwen4-Max's parameter count, benchmark scores and full-release timing have none of them been published.
TechNode's report adds the operational detail: multimodal input across text, image and video, tool calls during document and video processing, structured outputs, and the November 25 shutdown date for the preview build. Its 1-million-token context figure is the one number now traveling with the launch, and it is also the one to hold most loosely. No Alibaba model card states it for Qwen4-Max. The figure matches the published context length of the current flagship, Qwen3.8-Max. Separately, a Qwen Code pull request merged on October 28 (PR #15762, "size the qwen4 generation at 1M") set the client's context-window detector to size every model identifier beginning with "qwen4" at 1,000,000 tokens — a rule its own text says applies only to identifiers no upstream catalog describes, adopted at a time when no Qwen4 entry existed in the catalogs the client consults. A client setting does not by itself establish the served model's accepted context limit or its effective long-context performance, and it neither proves nor disproves TechNode's figure. What is established is narrower: as of November 11, the 1M number for Qwen4-Max rests on reporting, not on Alibaba documentation.
The missing list is longer: parameters, benchmarks, rate limits, per-token pricing, and any word on downloadable weights for Max. One part of the family story is documented, though. Qwen released Qwen3.8-Flash-Next as open weights on August 26, describing it as the first open-weight release on this generation's new architecture — so the architecture has a public, downloadable representative, even if the flagship does not.
How to use the window
Because the preview build disappears, the durable output of testing it is your own evaluation record. Run it against a fixed snapshot of a real repository, with the same prompts, tools and retry rules you use for your current model, and log the things aggregate scores hide: tool-call failures, edit acceptance, attempts needed per task, latency and token use. Preserve the model identifier, settings and timestamps with the results. You can measure the model on your own tasks without vendor benchmarks; what those measurements cannot do is predict the final release, whose specification has not been published.
Do not make a production dependency on the reported preview without confirming its lifecycle, limits and service terms. A preview with a published shutdown date and no data-handling terms in its announcement is a test bench, not a platform.
The open question
Qwen's current flagship line is documented in unusual detail — Qwen3.8-Max's parameter counts and context length are published figures. Qwen4-Max, for now, is the opposite: a name, a capability claim and a shutdown date. Whether the full release restores the documentation habit — and whether the open-weight line that began with Flash-Next extends to the rest of the family — is what the November 25 deadline will start to answer.
Sources
- [1] Official Qwen account (@Alibaba_Qwen), announcement post, November 11, 2026 — "Qwen4 is coming. Deeper knowledge. Richer world understanding. Stronger coding and agentic capabilities… Qwen4-Max is now in preview on Alibaba's Token Plan. You can also get early access to the new model on QwenWork, Qoder, Qwen APP, and QwenStudio." (read in full as reproduced by PC Watch)Read source
- [2] PC Watch (Impress), "「Qwen4」早くも登場、最上位「Qwen4-Max」を無料でも試せる", November 11, 2026 (free-plan trial; Token Plan event pricing from $3/month for new users; parameters, benchmarks and full-release timing all unpublished as of November 11; Qwen3.8-Flash-Next and Qwen3.8-Max background)Read source
- [3] TechNode China, "阿里上线Qwen4-Max预览版，支持百万Token上下文", November 11, 2026 (multimodal input, 1M-token figure, preview window November 11 10:00 – November 25 24:00 China time, preview taken offline afterwards)Read source
- [4] QwenLM/qwen-code, pull request #15762, "fix(core): size the qwen4 generation at 1M" (merged October 28, 2026; client context-window detector rule for qwen4 identifiers; catalogs consulted first)Read source