Nextbrowser
Blog
English

Qwen4-Max Is Out in Preview: What It Can Do and Where to Try It

Alibaba opened a preview of Qwen4-Max with a 1M-token context, image and video input and tool calling. Here is what the model can do and how to try it for free.

Kirill Zaikov avatar
Kirill ZaikovAuthorXLinkedIn
Qwen4-Max Is Out in Preview: What It Can Do and Where to Try It cover

Alibaba has opened a preview of Qwen4-Max, the first flagship model of the Qwen4 family. It has a 1-million-token context window, it reads images and video, and it can call tools and return JSON that follows a schema. The preview is short: it ends on November 25.

TL;DR: Qwen4-Max-Preview is available now. You can try it for free in Qwen Chat without an account, or use it through Alibaba's Token Plan API. The first independent benchmark puts it at 58.8/100: strong on visual tasks, weaker on reasoning games. In the upcoming Nextbrowser release, you will be able to test Qwen4-Max for free right inside the browser.

What we know about Qwen4-Max

According to Alibaba's Model Studio documentation, this is what the preview offers:

SpecQwen4-Max-Preview
Preview periodNovember 11 to November 25, 2026 (UTC+8). The model does not change during the preview and is retired after it
Context window1,000,000 tokens
Maximum output393,216 tokens, shared between thinking and the final answer
InputText, images and video
OutputText
ThinkingOn by default, with low, medium and xhigh effort levels (xhigh is the default)
ToolsFunction calling, web search and built-in tools
Structured outputJSON Object and JSON Schema
API formatsOpenAI Chat Completions, OpenAI Responses and Anthropic Messages

First benchmark results

Alibaba has not published an official benchmark table, but the model is already being tested independently. On StealthMark, Qwen4-Max-Preview scored 58.8/100 and took 12th place out of 58 models (measured on November 11):

StealthMarkQwen4-Max-PreviewQwen3.8-Max
Overall58.8 (12th of 58)62.4 (9th of 58)
Visual Intelligence66.363.9
General Intelligence54.661.6
Beach (SVG scene)78.4 (5th)76.9 (9th)
Pelican (SVG drawing)77.3 (7th)72.0 (19th)
Latch (reasoning game)67.3 (21st)82.6 (7th)

The picture is clear. The new model draws and builds visual scenes better than Qwen3.8-Max, but loses to it in games that test step-by-step reasoning. So the best way to try it is a visual task.

Still missing: the parameter count, a model card, open weights and a per-token price.

Where to try Qwen4-Max today

Qwen Chat (free, no account needed)

The simplest way is Qwen Chat:

  1. Open chat.qwen.ai. You can close the sign-in window and continue as a guest.
  2. Click the model name in the top-left corner and choose Qwen4-Max-Preview.
  3. Type your prompt or attach a file with the + button.
  4. Before the first answer, Qwen Chat asks you to confirm your age by choosing your birth year.

Choosing Qwen4-Max-Preview in the Qwen Chat model picker

Token Plan API

If you want to call the model from code or from a coding tool, it is available through Alibaba Cloud's Token Plan:

  • The model ID is qwen4-max-preview.
  • There is no pay-as-you-go option during the preview. Token Plan is a monthly subscription with credits; the personal Lite plan costs $6 per month, and new users pay $3 until November 30.
  • Because the API supports the OpenAI and Anthropic formats, you can plug it into tools like Claude Code, Cursor, Cline or Qwen Code.
  • The personal plan is meant for interactive use in coding tools and agents. Batch scripts and app backends are not allowed.

Qwen4-Max-Preview is also available in Alibaba's own apps, including QwenStudio, the Qwen app and Qoder.

A quick test: a 3D game in one prompt

Since the model is strongest on visual tasks, we asked it to build a small 3D game. Here is the prompt we sent in Qwen Chat:

text

The first attempt in the default Auto mode thought for about 25 minutes and then failed with "The request is ended!" without returning any code. In Fast mode, the same prompt worked. Along the way, the model ran its code interpreter dozens of times to check its work, and after several minutes it returned one HTML file with 646 lines of code.

The game, which the model named "Fox Run", ran on the first try:

Start screen of the game Qwen4-Max-Preview built

It did everything we asked and added things we did not: berries to collect, crates and logs as obstacles, sound with a mute button, a saved best score, a speed bar and swipe controls for phones. The scene is a low-poly road at sunset with pine trees, rocks and a low sun.

The fox running down the three-lane road

Our first run ended at 137 points, when the fox ran into an obstacle.

Game Over screen with the final score

Two practical notes. Qwen Chat cannot run the game itself because the site blocks external scripts like Three.js, so copy the code into a local .html file and open it in your browser. And for big generation tasks, switch the mode to Fast: the default mode may think for too long and time out.

Coming soon to Nextbrowser

In the upcoming Nextbrowser release, Qwen4-Max will appear in the model picker of the built-in AI chat, free to test. Download Nextbrowser or update to the latest version to get it as soon as it ships.

What comes next

The preview runs until November 25, and Alibaba has not announced a date for the full release. We are watching for an official benchmark table, the final API price and news on smaller Qwen4 models. In our next post, we will compare Claude Opus 5.5 and OpenAI's GPT-6 Astra.

FAQ

Is Qwen4-Max free?

You can try it for free in Qwen Chat, without an account. API access during the preview requires a Token Plan subscription. In the upcoming Nextbrowser release, you will also be able to test it for free inside the browser.

Does Qwen4-Max have open weights?

No. The preview is cloud-only, and Alibaba has not said whether weights will be released.

What happens after November 25?

Alibaba says the preview model will be retired when the preview period ends. It has not published a date for the full release.

How is Qwen4-Max different from Qwen3.8-Max?

Qwen3.8-Max is the current generally available flagship. On StealthMark, Qwen4-Max-Preview scores lower overall (58.8 vs 62.4) but higher on visual tasks (66.3 vs 63.9). It is weaker in reasoning games, where Qwen3.8-Max still leads.