After weeks of teasers, codenames and prediction-market speculation, the first piece of Alibaba’s Qwen 4 generation is finally something developers can touch. On November 11, 2026, Alibaba opened Qwen4-Max-Preview — a multimodal flagship model with a one-million-token context window — but only through its Token Plan subscription, only for two weeks, and with almost none of the documentation that usually comes with a flagship launch.
There are no published benchmarks, no parameter count, no per-token price and no downloadable weights. What Alibaba has published is a short specification sheet and a hard deadline: the preview runs from 10:00 on November 11 to 24:00 on November 25, 2026 (UTC+8), and the model will be retired when the window closes. It is a trial run, not a product launch — and the terms say as much.
What Qwen4-Max-Preview Is
Qwen4-Max-Preview is the first model from Alibaba’s Qwen4 generation that developers can call. According to QwenCloud’s official documentation, it is the next-generation flagship multimodal model in the Qwen family, built for coding, office productivity, front-end design and complex agentic tasks. It accepts text, images and video as input and returns text, with deep reasoning switched on by default.
The Qwen team says the model brings deeper knowledge, richer world understanding and stronger coding and agentic skills than Qwen3.8, the previous generation. Those are the team’s own claims, and with no benchmark table published at launch, they have not been checked by any outside evaluator yet.
The launch itself was characteristically low-key for a company that has made a habit of shipping first and announcing second. At 10:00 Beijing time on November 11, the official Qwen account posted simply “Qwen4 is coming,” and Alibaba listed the preview model on its Token Plan the same morning. Early access also runs inside Alibaba’s own apps — QwenWork, Qoder, the Qwen app and QwenStudio.
A 1M-Token Context and a 393,216-Token Output Budget
The headline numbers are generous. Qwen4-Max-Preview ships with a one-million-token context window and a maximum output of 393,216 tokens — but that output ceiling is shared between the model’s internal reasoning and its final answer, not reserved for the answer alone.
That shared budget matters in practice. Thinking is enabled by default, controlled by a reasoning_effort parameter with three settings — low, medium or xhigh, with xhigh as the default — and the documentation warns developers to leave room for reasoning when setting output limits, since the thinking tokens count against the same 393,216-token ceiling. The docs also note that reasoning_effort and thinking_budget must not be set together.
On the input side, the model handles text, images and video, with screenshot analysis, document understanding and video analysis named as intended uses. It supports function calling, web search and built-in tools, structured output in JSON Object and JSON Schema formats, and it speaks the OpenAI Chat Completions, OpenAI Responses and Anthropic Messages API formats — so it can sit behind coding agents such as Qwen Code, Claude Code, Cline or Codex with a configuration change rather than a rewrite.
A million-token context puts Qwen4-Max-Preview in the same conversation as the biggest frontier models of the moment, including Google’s Gemini 4 Carbon, which reached general availability this week. The difference is what comes with the number: Gemini 4 Carbon arrived with benchmark tables and pricing, while Qwen4-Max-Preview arrived with a spec sheet and a countdown clock.
Token Plan Only — and a Hard Stop on November 25
Access is the most unusual part of this launch. Qwen4-Max-Preview is available only through Alibaba’s Token Plan subscription service. Pay-as-you-go API access is explicitly not supported during the preview, an ordinary Model Studio key does not open this model, and requests go to a dedicated base URL — https://token-plan.maas.qwencloudapi.com — with the model ID set to qwen4-max-preview.
The window is fixed: 10:00 on November 11 to 24:00 on November 25, 2026, UTC+8. Alibaba says the model will not be changed during that period and will be retired when it ends. There is no announced extension and no migration path to a successor model, so anything built on the qwen4-max-preview model ID will simply stop working after November 25.
That deliberate scarcity looks like a controlled experiment. Alibaba gets two weeks of real developer usage and telemetry — which prompts, tools and workloads the model actually handles — before committing to a full production release. Developers get an early look at where the next Max model is heading, inside the coding tools they already use. The trade is clear, and it is one-sided by design: you can test it, but you cannot build a business on it yet.
What Alibaba Didn’t Publish — and Why It Matters
A flagship model launch with no benchmarks is rare, and the gaps in the Qwen4-Max-Preview documentation are worth listing plainly:
- No benchmark scores, vendor-reported or independent, on reasoning, coding or any other evaluation.
- No parameter count and no model card describing capabilities, limits or training data.
- No per-token price — the Token Plan subscription is the only route in, and its terms are separate from the model itself.
- No downloadable weights and no license terms, so there is nothing to self-host or fine-tune.
That absence is also a warning about the numbers already circulating online. Charts and scores attributed to Qwen4-Max — including a figure said to come from Artificial Analysis and an 82% SWE-bench claim for a “Qwen4-Coder” — do not trace back to any primary source. Artificial Analysis had no Qwen4-Max score on its leaderboard on launch day, and Alibaba has announced no Qwen 4 Coder at all; the current coding-focused models are the Qwen3-Coder line. Treat any Qwen4-Max benchmark you see this week as a third-party test of the preview or an unverified leak, not as an official result.
The same caution applies to the rumors that preceded the launch. An unreleased model codenamed Kalmia has been appearing in anonymous Arena testing since late October, and some testers have attributed it to Qwen 4, but neither Alibaba nor the Qwen team has confirmed any connection. A stealth model called Qoduck appeared inside Alibaba’s Qoder coding IDE on November 7, again with no confirmation of what it is. The preview that actually shipped is the first confirmed fact in weeks of speculation.
The Architecture Behind the Preview
While the preview model itself comes with no architecture details, part of the Qwen 4 architecture is already public — inside a downloadable model. On August 28, 2026, Alibaba released Qwen3.8-Flash-Next and described it in its model card as “an experimental preview of the architecture that will underpin Qwen4,” shipped early so developers could prepare.
Flash-Next is a sparse mixture-of-experts model with 125 billion total parameters but only about 6 billion active per token, plus a 51-billion-parameter n-gram embedding component that trades transformer computation for memorized lookup, native image and video input, and a 262,000-token context extensible to one million. Its open weights are available under the qwen-community-1.0 license. Whether the flagship Qwen 4 scales those choices up or refines them is unknown, but the direction is on the record: leaner activation, a memory layer, and multimodality as the default rather than a variant.
The roadmap around it was set out at Alibaba’s Apsara Conference on September 22, when Qwen project lead Liu Dayiheng said Qwen 4 was in training on a new-generation architecture and would be released “very soon,” and that Qwen 4.5 and Qwen 5 would scale to 5 to 10 trillion parameters. Alibaba has named four Qwen 4 tiers — Max as the flagship, Plus as the balanced multimodal tier, Flash for low-latency, high-volume work, and a 27B open-weights model for local deployment. Only Max has shipped, and only as a preview. The other three have no specs, pricing or dates.
The staged pattern is familiar from the Qwen 3.8 generation, which moved from preview to flagship to open weights in about a month. If Qwen 4 follows the same cadence, the two-week Max preview ending November 25 is likely the opening move rather than the main event — though Alibaba has said nothing about what follows it.
How to Try It
Trying Qwen4-Max-Preview takes three steps, all spelled out in the official documentation. First, subscribe to Alibaba’s Token Plan (Individual or Team) and obtain a dedicated API key. Second, point your client at the Token Plan base URL, https://token-plan.maas.qwencloudapi.com. Third, set the model ID to qwen4-max-preview.
The endpoint works with Qwen Code, Cline, Claude Code, Cursor, OpenCode and Codex, and early access is also live in QwenWork, Qoder, the Qwen app and QwenStudio. For multi-turn conversations, the docs advise returning all historical thinking in the assistant messages’ reasoning_content field rather than concatenating it into the content.
The practical advice is simple: use the window to test, not to ship. Evaluate the model against your own coding and agent workloads, keep your serving behind an OpenAI-compatible interface so swapping models later is a one-line change, and do not wire anything production-critical to a model ID with a published retirement date two weeks out.
What This Means for the AI Race
Qwen4-Max-Preview arrives in one of the busiest model-release weeks of the year. The same week brought Google’s Gemini 4 Carbon into general availability, Meituan’s open-weight LongCat-2.5 agent model, DeepMind’s AlphaGenome 2 for DNA research, and Meta’s tiny, cheap Muse Ember model for agent work. Against that field, Alibaba’s move stands out less for what it shipped than for how it shipped it: a flagship preview with no scoreboard attached.
That restraint may be the point. With no benchmarks to argue about, the preview forces the evaluation onto real usage — developers running their own tasks through their own tools for two weeks. If the model is genuinely strong at coding and agentic work, that evidence will surface in developer reports long before any official table appears. If it is not, the same channel will say so, and Alibaba will have learned it before committing to a general release.
The open question is what happens on November 26. A full Qwen4-Max launch with pricing and benchmarks, a Plus or Flash tier, or open weights for the 27B model would all fit the pattern Alibaba followed with Qwen 3.8. Until then, Qwen4-Max-Preview is best understood as a two-week audition — a chance to see the next Qwen flagship early, on Alibaba’s terms, with the clock already running.
Frequently Asked Questions
What is Qwen4-Max-Preview?
Qwen4-Max-Preview is Alibaba’s first Qwen4 model available to developers, launched November 11, 2026. It is a multimodal flagship model with a 1M-token context, built for coding, office work, front-end design and complex agent tasks.
How long is the Qwen4-Max preview available?
From 10:00 on November 11 to 24:00 on November 25, 2026 (UTC+8). Alibaba says the preview model will not change during that window and will be retired when it ends.
How do I access Qwen4-Max-Preview?
Only through Alibaba’s Token Plan subscription, using a dedicated API key and the base URL https://token-plan.maas.qwencloudapi.com with the model ID qwen4-max-preview. Pay-as-you-go API access is not supported.
Are there benchmarks for Qwen4-Max?
Not from Alibaba. The preview launched with no benchmark table, parameter count, model card or per-token price, so any Qwen4-Max scores circulating online are unverified third-party tests or leaks.
Is Qwen4-Max open source?
No weights or license have been published for Qwen4-Max-Preview. Alibaba has announced a 27B open-weights tier for the Qwen 4 family, but it has no specs, pricing or release date yet.
When will the full Qwen 4 release arrive?
Alibaba has not given a date for the full release or for the Plus, Flash and 27B tiers. Its launch message said only “Qwen4 is coming.” The Qwen 3.8 generation moved from preview to flagship in about a month, which suggests the wait may not be long.
The Bottom Line
Qwen4-Max-Preview is a flagship you can use but cannot rely on — a 1M-token-context, multimodal reasoning model, available for exactly fourteen days, through one subscription, with no benchmarks, no price and no weights. That combination is frustrating for anyone trying to rank it against the frontier models it is meant to challenge, and it is deliberate: Alibaba is gathering real-world evidence before it commits to a general release.
For developers, the window is worth using. Two weeks is enough to learn whether Qwen4-Max earns a place in a coding or agent workflow, and the OpenAI-compatible endpoint makes the test cheap to run. Just do not build on it. The model ID retires on November 25, and whatever Qwen 4 becomes next — a priced Max, a Plus tier, an open 27B — is a story for after the preview ends.
Sources: QwenCloud’s official Qwen4-Max-Preview documentation, Emergent’s report on the launch, and YottaLabs’ Qwen 4 release tracker.
