Chinese Models Desk
Chinese Models Desk

Alibaba Opens Qwen4-Max Preview: 1M Tokens, Two Weeks, and No Benchmarks

Alibaba's Qwen4-Max preview offers a 1M-token window and video input through Token Plan until November 25, but ships without benchmarks, pricing, weights or a license.

John EditorNov 13, 2026 10m read
ShareWhatsAppXFacebook

For months, Alibaba's next-generation Qwen family has existed mostly as a slide, a rumor and a line of code in someone else's pull request. On November 11, 2026, that changed, at least in part. Alibaba opened a preview of Qwen4-Max, the flagship of the new family, to paying developers and to users of its own apps. It arrives with a 1,000,000-token context window, image and video input, and a hard expiry date: November 25, 2026. It also arrives with no benchmarks, no parameter count and no weights.

This piece, written from the analytical side of the desk, separates what Alibaba has actually published from what the market is projecting onto it. The gap between the two is the real story.

What Alibaba Actually Launched

The announcement was brief. As TechNode reported on November 11↗, the model supports image and video inputs, handles up to a million tokens of context, and is reachable through Alibaba's Token Plan subscription. Early access also runs inside the Qwen app, Qoder, QwenWork and QwenStudio.

The first written specification comes from Alibaba's own Qwen Cloud model changelog↗, which describes Qwen4-Max-Preview as "the next-generation flagship multimodal model in the Qwen family," combining text generation, image and video understanding, and deep reasoning. The stated target workloads are coding, office productivity, frontend design and complex agentic tasks. The entry is explicit on access: "Available through Token Plan only."

The published specifications

Pulling together the changelog and the more detailed write-ups from Gadget Pilipinas↗ and Fello AI↗, the confirmed technical surface looks like this:

  • Context window: 1,000,000 tokens, with a maximum output of 393,216 tokens that is shared between the model's thinking and its final answer.
  • Modalities: text, image and video input, with deep reasoning switched on by default and a default effort level of xhigh; developers can lower it to low or medium through a `reasoning_effort` parameter.
  • Tooling: function calling, web search and built-in tools, plus structured outputs in both JSON Object and JSON Schema modes.
  • API compatibility: OpenAI Chat Completions and Responses formats, and the Anthropic Messages format, so existing agent tooling can be pointed at it without a rewrite.
  • Identifiers: the model ID is `qwen4-max-preview`, with `qwen4.0-max-preview` as an alias, and billing records show the alias.
"Qwen4 is coming. Deeper knowledge. Richer world understanding. Stronger coding and agentic capabilities." > — Alibaba's Qwen account, November 11, 2026, as quoted by Fello AI

That is a recognizably agent-first positioning. The 393K output ceiling and the xhigh default tell you what Alibaba expects customers to do with it: long, tool-heavy sessions in which the model thinks for a while and then writes a lot.

A Fifteen-Day Window, and the Fine Print

The preview opens on November 11 and, per Alibaba's documentation as relayed by Gadget Pilipinas, the model is retired once it closes on November 25. Fello AI quotes the documentation promising that "no changes will be made during this period," and notes that Alibaba has not said what replaces the preview afterwards.

That matters for anyone thinking about production use. A model that disappears after two weeks is an evaluation vehicle, not a dependency. If you wire it into a workflow, you need a fallback, such as the generally available Qwen3.8 line, already configured.

Pricing is a credit system, not a price list

There is no per-token price. Token Plan bills in credits drawn from a monthly quota. According to Alibaba's Token Plan pricing page↗ and the coverage above, individual tiers run from $6 a month (Lite) to $43 a month (Pro), and during the preview the model draws credits at as little as 20% of the standard rate. Calls between 22:00 and 08:00 Beijing time fall to 8%.

Two consequences follow:

  • Because Alibaba has not published the standard credit rate, nobody can yet convert the promotional discount into a dollars-per-million-tokens figure, and any such figure you see this week is an estimate.
  • Pay-as-you-go access is not supported, so an ordinary Qwen Cloud API key will not reach the model; you need an active Token Plan subscription.

The subscription-only route is a notable commercial choice. Token Plan is pitched at coding assistants and agent tools, so Alibaba is steering its newest flagship toward the developer-seat model that Western labs have used for their coding products, rather than toward metered API revenue.

What Is Missing

The omissions are as informative as the specifications. As of this writing, Alibaba has published:

  • No benchmark results of any kind for Qwen4-Max, so claims about how it ranks against Western frontier models cannot be verified.
  • No parameter count or architecture description, so the size of the model and whether it uses a sparse design are unconfirmed.
  • No weights on Hugging Face, no model card and no license, which leaves open whether this tier will follow the open-release path.
  • No dates for the other tiers that have been shown on slides: Qwen4-Plus, Qwen4-Flash and Qwen4-27B.

I would treat any Qwen4-Max benchmark table circulating in the next two weeks as unofficial unless it links back to an Alibaba page. Fello AI makes the same point, and it is the correct discipline here.

The Teaser Campaign and the Slipped Date

Alibaba built anticipation before shipping. On November 6, according to Gadget Pilipinas, the Qwen team posted a poster of birds forming the shape of a fish over a dark sea, with a classical Chinese line that stops one character short and the words "The becoming begins." The missing character, as Asia Times noted in the same report, is 鲲 (Kun), the giant fish of the Zhuangzi that becomes a bird. It is a fitting image for a family that has been transforming from an open-weight workhorse into a tiered commercial platform.

The timing also slipped slightly. Fello AI recounts that an attendee at Alibaba Cloud's Amsterdam Apsara event on October 29 reported the company's CTO saying "Qwen4 next week," which implied November 2 to 8. The preview landed on the following Wednesday. That is a modest miss, but it is a reminder that the only on-the-record statement before November 11 was a roadmap line from the September Apsara Conference saying Qwen 4 was "in training."

The Architecture You Can Already Inspect

Although Qwen4-Max itself is opaque, Alibaba has said that part of the design is already public. According to Fello AI's reading of the team's GitHub notes, Qwen3.8-Flash-Next, released on August 26, 2026, "serves as an early preview of the architecture used in Qwen4," playing the same role that Qwen3-Next played for the Qwen 3.5 generation. The model card figures cited there are:

  • 125B total parameters with 6B active per token, plus a 51B n-gram embedding table and a 4B multi-token-prediction head.
  • A 262,144-token native context, extensible to 1,000,000 tokens.
  • Gated DeltaNet attention paired with Qwen Sparse Attention.
  • A Qwen Community License 1.0, which requires a separate license for businesses that sell model access through an API or build coding and office assistants on it.

The design bet is straightforward: park a large share of parameters in a cheap lookup table that can sit in ordinary memory, and let only a few billion parameters do the work on each token. If Qwen4 keeps that shape, its models should be cheaper to serve than their total size implies. I want to be careful, though. Alibaba's Qwen4-Max-Preview page does not describe the model's internals, so the link between Flash-Next and Qwen4-Max is the team's stated direction, not a confirmed specification of the flagship.

How This Fits the Competitive Picture

Context length is no longer a differentiator among Chinese labs. The same Qwen Cloud changelog↗ shows that DeepSeek-V4.1-Flash-1017, a 552B-parameter mixture-of-experts model, already offers a 1M-token window with a 384K maximum output, and that GLM-5.3 from Zhipu advertises a 1M window with 128K output. Alibaba's own Qwen3.8-Max and Qwen3.8-Flash carry native 1M windows too.

So the million-token headline matches the current Chinese frontier rather than extending it. What Qwen4-Max adds, on paper, is a larger output ceiling, a default of maximum reasoning effort, and multimodal input folded into a flagship rather than a side model. Whether that translates into better results is exactly what the missing benchmarks would settle.

Open weights are the open question

Alibaba's recent history complicates any prediction. Qwen3.8-Max launched through the API first, and its weights followed as Qwen3.8-2.4T-A95B, which the changelog lists as the open-source version of the flagship series: 2.4 trillion total parameters, roughly 95 billion active, a 1M context window. The licensing in that generation varied by tier, from Apache 2.0 for the 27B model to custom terms for the largest. Fello AI↗ also reports that a Manifold market on open Qwen 4 weights by November 25 fell from 83% to 59% between November 5 and November 11, though that market is small and should be read as sentiment, not insider knowledge.

My read is that the subscription-only, time-limited preview is consistent with either path. It could be a conventional prelaunch evaluation ahead of a wider release, or it could signal that the top tier is being kept closer to the commercial platform than earlier flagships were. Nothing published so far separates those readings.

What Developers Should Do Now

For practitioners, the sensible response is measured:

  • Test it on your own tasks in QwenStudio or through a Token Plan subscription before November 25, especially long-context coding and document workloads where the 1M window and 393K output matter.
  • Do not plan a product around it: there is no price, no license and no committed successor, and the model is scheduled for retirement.
  • Keep Qwen3.8 as the stable baseline, since it is generally available through standard Qwen Cloud access with published context windows and documented endpoints.
  • Watch three places on and after November 25: the Qwen Cloud changelog, the Qwen account's announcements and the Qwen organization's Hugging Face page for any new repository.
The date to watch is November 25. That is when Alibaba's preview window closes, and when the company will have to say what comes next.

Analysis: A Preview That Sells Access, Not Proof

Stepping back, the Qwen4-Max launch is more a statement about distribution than about capability. Alibaba chose to give paying developers hands-on time, lock the model behind a subscription, discount it aggressively for night-time use in China, and withhold the evidence that would let outsiders rank it. That is a confident posture for a lab whose model family counts among the most downloaded in the world, but it asks the market to take performance on trust for now.

The comparison with Western launches is instructive. Frontier labs in the United States typically pair a flagship release with benchmark tables, system cards and clear per-token pricing. Alibaba has done the opposite for this preview: published the interface, the limits and the dates, and left the quality claims to user experience. For developers the practical consequence is simple. The next fourteen days are an opportunity to form your own view, and the days after November 25 will show whether Qwen4 becomes an open ecosystem again or a more tightly held commercial product.

Either way, the Qwen family has clearly entered a new phase, with a flagship that arrives as a measured, metered, time-boxed experiment rather than a drop of weights onto Hugging Face. I will revisit the story when Alibaba publishes benchmarks, pricing or a license, any of which would change the assessment above.

#Qwen4#Alibaba#Qwen4-Max#Token Plan#Chinese AI

Comments

Open discussion — no account needed. Be respectful.

0/4000
Loading comments…