Alibaba has opened early access to Qwen4-Max, the flagship of its next-generation Qwen family – but only for 15 days: the preview will be switched off after November 25. During the preview the model costs up to 80% less than the standard rate, and up to 92% less at night, Beijing time.
You can try the model in the Qwen app, or plug it into your own tools with a Token Plan subscription starting at $3 a month. The company has not published any benchmark results yet. Here is what we know about the first Qwen4 model, where to try it, what limits come with the subscription and when to expect the full release.
Key facts
- What: Qwen4-Max-Preview is the first model of the new Qwen generation. It understands text, images and video and reasons before it answers. The API model ID is
qwen4-max-preview. - Built for: complex coding, front-end work, office tasks and agentic workflows, according to Alibaba.
- Context: 1M tokens, with up to 393,216 output tokens including reasoning.
- Where: the Token Plan subscription (from $3 a month for new users) and early access in QwenWork, Qoder, the Qwen app and QwenStudio.
- How long: until 24:00 on November 25, Beijing time (16:00 UTC), after which the model will be retired.
- What’s missing: benchmarks, a model card, a parameter count, a per-token price and weights.
- A possible early version: from November 7 to 11, an anonymous model called Qoduck ran in Qoder. Its tokenizer matches Qwen, and it was switched off the minute the preview went live. Alibaba has not confirmed the link.
Alibaba’s official release
The Qwen team announced the preview on November 11 at 02:01 UTC with a post titled “Qwen4 is coming”. By then the model was already live: according to the documentation, the preview opened on Token Plan at 10:00 Beijing time, or 02:00 UTC. Early access in QwenWork, Qoder, the Qwen app and QwenStudio followed the same day. Alibaba promises deeper knowledge, richer world understanding and stronger coding and agentic skills. The model is aimed at real-world work: complex code, knowledge work, front-end builds and professional software.
The launch amounted to a short social post (on X and on China’s Weibo) and two documentation pages. There is no technical report, model card or benchmark table – the materials Alibaba usually ships with its Max models.
Alibaba Cloud’s documentation describes Qwen4-Max-Preview as the next-generation flagship multimodal model in the Qwen family. It can call functions, search the web, use built-in tools and return output that follows a JSON Schema. The model is compatible with the OpenAI APIs (Chat Completions and Responses) and with Anthropic Messages, so it can be connected to Claude Code, Cursor and other coding agents with a dedicated endpoint and key.
Specs
| Spec | Qwen4-Max-Preview |
|---|---|
| Model ID | qwen4-max-preview (alias qwen4.0-max-preview) |
| Input | text, images, video |
| Context | 1M tokens |
| Max output | 393,216 tokens, including reasoning |
| Reasoning | on by default; low, medium and xhigh levels (xhigh by default) |
| Tools | function calling, web search, built-in tools |
| Compatibility | OpenAI Chat Completions and Responses, Anthropic Messages |
| Access | Token Plan subscription only, no pay-as-you-go |
| Preview window | 10:00 on November 11 to 24:00 on November 25, Beijing time (02:00 UTC on November 11 to 16:00 UTC on November 25) |
Developers should watch for one detail that can break existing code. Starting with this model, max_tokens in Chat Completions caps reasoning and the answer together, not just the answer. A limit tuned for earlier Qwen models may truncate output: part of the budget goes to reasoning, and by default the model reasons at the highest level, xhigh. In Qwen Code, the default output limit for this model is 32,768 tokens, so longer answers require raising it manually.
Where to try it
Besides Token Plan, Alibaba opened early access in four of its own products: QwenWork (office agents), Qoder (an AI code editor), the Qwen app and QwenStudio. TestingCatalog also spotted the model in Qwen Chat, including Agent Mode. To connect it to your own tools over the API, you need Token Plan.
What Token Plan costs
Token Plan is Alibaba’s paid subscription. Instead of paying per token, you spend Credits, a single currency for every model in the plan. Personal plan prices:
| Plan | Price per month | New users | Credits per month | Concurrent agents |
|---|---|---|---|---|
| Lite | $6 | $3 | 20,000 | 1–2 |
| Essential | $10 | $5 | 45,000 | 2–3 |
| Standard | $18 | – | 80,000 | 3–4 |
| Pro | $43 | – | 200,000 | 6–8 |
The new-user price is valid until 23:59 on November 30, Beijing time, and only for accounts that have never bought a personal Token Plan. If you run out of Credits, you can buy an extra bundle of 20,000 Credits for $15.
Credit consumption depends on the model, the number of tokens, the reasoning level and the tools called, and exact figures only appear in the console after the fact. That means Qwen4-Max cannot yet be compared with Claude or GPT on price per token.
During the preview, Qwen4-Max consumes Credits at as little as 20% of the standard rate, so the quota lasts up to 5 times longer. From 22:00 to 08:00 Beijing time (14:00 to 00:00 UTC), the rate drops to 8%.
Token Plan limits
- Interactive use only. The plan is intended for use inside tools such as Claude Code, Cursor, Qwen Code, Qoder and OpenClaw. Scripts, custom backends and batch calls are prohibited, and violations can lead to a suspended subscription or a banned key. Strictly speaking, this means you cannot run the model through your own test suite automatically.
- Separate key and endpoint. You need a Token Plan key and the
token-plan.ap-southeast-1.maas.aliyuncs.comendpoint – a regular Model Studio key will not work. - Singapore and data transfer. The plan can only be bought in the Singapore region, and the model runs in Global mode: prompts and outputs may be processed in other countries. Alibaba leaves compliance with data laws to the user.
- No cancellation. A personal subscription cannot be cancelled after purchase.
- One per person. Keys cannot be shared, and each verified identity gets one subscription per edition.
What Alibaba hasn’t disclosed
Qwen4-Max has neither a benchmark table nor a model card, so every claim about its quality comes from Alibaba itself, AlphaSignal notes. The parameter count, license and per-token price are unknown, and the company has said nothing about weights. Social media posts already claim Qwen4-Max beats GPT-6 and Claude Fable 5, but nothing backs that up.
Qoduck: the anonymous model that was likely Qwen4-Max
Four days before the announcement, a model with no named developer appeared in Qoder, Alibaba’s AI code editor. On November 7, Qoder added Qoduck to its model picker with the line “Don’t ask where it came from.” On its promotion page, Qoder calls it a stealth model whose identity has yet to be disclosed and promises details “at an appropriate time.” Users got a free daily allowance, even at the highest reasoning level.
The promotion ended on November 11 at 10:00 Singapore time, and the model was removed from Qoder. That same minute (Singapore and Beijing share a time zone), the Qwen4-Max preview went live, and Qoder was on the list of products with early access to it.
Why Qoduck is linked to Qwen
- Tokenizer. The AI Tracker Bot account compared how Qoduck splits text into tokens and found a match with the Qwen family. On its knowledge check, Qoduck beat every released Qwen model, which led the tester to conclude it was a next-generation model rather than Qwen3.8.
- Teaser. The model appeared a day after the Qwen account posted a cryptic teaser reading “The becoming begins.”
- Price. Qoduck’s standard rate in Qoder matched Qwen3.8-Max, according to Stealth Models, so some users concluded it was a Max model. Others bet on Qwen4-27B or Flash.
- Chinese media. The outlet 53AI flatly stated that Qoduck is Qwen4-Max-Preview, but offered no evidence.
Neither Alibaba nor Qoder has confirmed the link. But the coincidences pile up: a Qwen tokenizer, Alibaba’s own product, results above every released Qwen, and a shutdown at the exact minute Qwen4-Max launched. That is why we believe Qoduck was Qwen4-Max or an early build of it. This is our assessment, not a fact: the preview may run a different, later version of the model.
What Qoduck showed
Qoduck has no benchmarks, only hands-on tests and demos. The most detailed test was run by Chinese blogger techgogogo. He gave the model six tasks at the highest reasoning level and called it strong, but stopped short of deciding whether it was Qwen4. The best results were an SVG of a pelican riding a bicycle, which the author called a rare top-tier result, and a Trello-style kanban board without a single bug. The misses were minor: the fletching on the arrow in the bow drawing was reversed, and a mini-game in a browser-based operating system could only be controlled with the mouse.
Qoder itself collected 11 user builds from the first 48 hours: a voxel White House with 68 walkable rooms, a floating palace in a single HTML file and a QR-code Chrome extension shipped the same day. In the same thread, the company compared Qoduck with DeepSeek V4.1 Flash on one task: 40 minutes versus 70. That is Qoder’s own marketing, not an independent measurement.
Reactions on X were mostly positive. One tester says Qoduck outperformed Qwen3.8-Max, another preferred it to Claude Opus 5.5 on the same prompt. But these are one-off comparisons with no methodology. 53AI describes the model more cautiously: it responds fast and writes simple code cleanly, but its quality is uneven on multi-step tasks and when editing an existing project, and its code completion is underwhelming for a Max-tier model.
If Qoduck is Qwen4-Max, the picture is this: the model is strong at building an interface, a game or a drawing from scratch in a single file, and weaker at long sessions with an existing codebase – the very work people connect models to agents for.
A chart showing Qwen4-Max scoring 53 on the Artificial Analysis index is also circulating. Stealth Models could not trace its source: the model does not appear in Artificial Analysis’s rankings or posts.
When to expect the full release: the Qwen3.8-Max precedent
Alibaba ran a similar playbook this summer. On July 19, Qwen3.8-Max-Preview launched first on Token Plan, Qoder and QoderWork, with a 90% daytime discount on Credits. The full Qwen3.8-Max arrived on August 3, 15 days later: 2.4 trillion parameters in a MoE architecture and a 1M-token context. Its API costs $2 per million input tokens and $6 per million output tokens. For comparison, Claude Opus 5.5 costs $4 and $20, and Claude Fable 5.2 costs $8 and $40.
If Alibaba repeats the pattern, the full Qwen4-Max should arrive in late November or early December. That is our estimate: the company has not given a date.
What else is coming in the Qwen4 family
Alibaba unveiled the Qwen4 lineup on September 22 at its Apsara conference in Hangzhou: the flagship Max, the fast Flash, the balanced Plus and the open Qwen4-27B. Max shipped first, and there are no dates for the other three. For anyone running models on their own GPU, 27B is the one to watch: it is the only one Alibaba calls open.
What this means in practice
- Test it on your own tasks. If you work in Claude Code, Cursor or Qwen Code, the $3 Lite plan is the cheapest way to compare Qwen4-Max with your current model before November 25. Give it edits in an existing project rather than from-scratch demos: according to the Qoduck reviews, that is where it struggles. Heavy jobs are cheaper between 14:00 and 00:00 UTC.
- Don’t build production on the preview. The model will be retired after November 25, there is no pay-as-you-go option, and using the plan in your own backend is prohibited.
- Check
max_tokens. Old limits may truncate answers, because they now include reasoning. - Wait for independent tests. For now, only Alibaba vouches for the model’s quality, and there is no way to verify rumors that it beats GPT-6 and Fable 5. More news on LLMs is in its own section.
