Qwen4-Max Preview review: Coding tests, GameFeel scores, pricing and setup

Qwen4-Max Preview is an early release of Qwen’s flagship Qwen4 model for coding, office work and multistep agent tasks.

It accepts text, images and video and supports web search, tool calling and structured output. Its 1-million-token context window can accommodate large amounts of code, documents and screenshots in the same task.

Official Qwen4-Max Preview documentation

Qwen4-Max Preview specifications

The API preview runs from 10:00 on November 11 until the end of November 25, Beijing time. This preview model will then be retired.

Qoder offered an anonymous test model called Qoduck for four days before replacing it with Qwen4-Max Preview in the model picker. Check the dates on earlier demos, which may have used a Qoduck build.

Qoder launch announcement

As of November 11, 2026, Qwen4-Max Preview had no entry on the public Artificial Analysis, Arena Text or Arena WebDev leaderboards. The official launch materials did not provide a full set of benchmark results.

One independent result comes from GameFeel Benchmark.

GameFeel RE: 10.3, below Qwen3.8-Max

GameFeel measures abilities related to game experience. RE (Reverse Evaluation) is one of its seven categories and tests how well models assess proposed game changes, especially their effect on balance.

Qwen4-Max Preview trailed Qwen3.8-Max and several early Qwen4 models on RE. That one result cannot establish how the preview compares with its predecessor at code generation overall.

Original GameFeel results post · Score breakdown · Benchmark methodology

The code people shared on launch day offers a closer look at what the model can build and where it struggles.

After the preview opened, developers posted websites, small games and SVG animations on X. Some outputs ran on the first attempt. Animation details and gameplay still varied across the examples.

1. An interactive 3D lighthouse island from one prompt

Developer AJ tried the same 3D lighthouse island prompt with several models.

Qwen4-Max Preview generated an interactive scene in one attempt and added working controls that the original prompt had not requested.

AJ also reported that Qwen4-Max Preview used fewer tokens than DeepSeek V4.1 Flash on his version of the task.

The result shows what the model can do with a standalone interactive page. Token use on one task offers a limited basis for comparing coding costs.

Original 3D lighthouse island test

2. A pelican on a bicycle, with improved SVG animation and rough details

Developer Leo tried the pelican-on-a-bicycle SVG animation again in Qoder.

He saw some improvement over the earlier Qoduck build, although he found the difference small.

Another tester pointed out that the pelican’s feet did not line up with the bicycle pedals.

An AI-generated animation can look complete while still getting the coordination between moving parts wrong.

Leo’s SVG animation test · Original post showing the pedal detail

3. Crossy Road, with results close to Claude Opus 5.5

Developer Mark Santos gave Qwen4-Max Preview and Claude Opus 5.5 Medium the same prompt.

build the crossy road game

Both models produced an HTML game. Mark downloaded and compared them and gave Claude Opus 5.5 a slight edge.

Original Crossy Road comparison

These examples show that Qwen4-Max Preview can build standalone websites, 3D scenes and small games. The quality varied by task.

Most public examples involve building a standalone frontend in one attempt. Large repository changes, difficult debugging and long-running agent workflows need more testing.

There are three main ways to try it: Qwen Chat, Qoder and API access through QwenCloud Token Plan.

1. Try it for free in Qwen Chat

The quickest way to try it is Qwen Chat.

Sign in and choose Qwen4-Max Preview in the model picker. To test a website or a small game, switch to Agent mode and select the model there.

A free account was reported to have used the model on launch day. Check your account for the available allowance.

Free account test · Qwen launch announcement

To test its coding ability yourself, try this prompt.

Open the result and try the filters, month switcher and reset button. Then narrow the browser window to check the layout.

2. Use Qoder in your development environment

Use Qoder to let the model read and edit files in a project.

Start a new task, click the current model name at the bottom of the composer and choose Qwen4-Max-Preview from the list.

Qoder’s launch announcement said the preview was open across its plans. The Credits rate can vary, so check the rate in the model picker before you send a task.

Qoder model selector guide

3. Set up API access through QwenCloud Token Plan

Use QwenCloud Token Plan to access Qwen4-Max Preview in developer tools that support the OpenAI or Anthropic protocols.

Here is how to set it up in Qwen Code Desktop.

1. Subscribe to Token Plan in the QwenCloud console and get a plan-specific API key.
2. Go to More → Settings → Models → Add Model.
3. Select QwenCloud → Token Plan → Singapore, enter the API key and choose qwen4-max-preview.

QwenCloud’s international service provides the following compatible endpoints.

Token Plan Individual is for interactive use in coding and agent tools. Its terms exclude application backends, batch processing and automation scripts.

Qwen Code setup guide · Token Plan quickstart · Individual plan terms

You can try Qwen4-Max Preview for free in Qwen Chat. API access requires a QwenCloud Token Plan subscription. Pay-as-you-go token billing is unavailable during the preview.

Here are the individual plan prices on QwenCloud’s international service.

Official QwenCloud Token Plan pricing

There are promotional Credits rates during the preview, and the published rates vary by platform.

QwenCloud’s international documentation lists a promotional rate as low as 20% of standard Credits, or as low as 8% from 22:00 to 08:00 Beijing time. Alibaba Cloud also announced an overnight offer.

Qoder’s launch post mentions a 0.1× Credits rate. Its model documentation lists 0.5×.

Check the Credits rate where you plan to use the model. Different platforms deduct Credits at different rates.

Alibaba Cloud offer announcement · Qoder launch offer

As of November 11, 2026, Qwen4-Max Preview’s model weights were unavailable, so you could not download it for local deployment.

Qwen also had not announced official release dates or weight downloads for Qwen4-27B, Qwen4-Flash or Qwen4-Plus.

Qwen model release notes

The claim that Qwen4 outscored Claude Opus 5.5 on gaming benchmarks refers to a specific early model.

Qwen4-Flash Early averaged 49.9 across GameFeel’s seven categories, compared with 47.2 for Claude Opus 5.5 XHigh. Those scores came from the early Flash build.

Original GameFeel test post

For a related model with published weights, see Qwen3.8-Flash-Next. The Qwen team describes it as an experimental preview of the architecture behind Qwen4.

It is worth trying for websites, interactive pages and small games. The launch-day examples include an interactive 3D scene from a single prompt, and you can try it in Qwen Chat.

For production development, those examples leave open how the model compares with the previous Qwen generation or competitors such as Claude.

The 10.3 in GameFeel RE shows a weak result on that task. The public frontend demos show what it can generate. Broader benchmarks and project tests would help assess both capabilities.

The API preview ends on November 25, 2026. Teams using it for daily development should plan how to switch models once the preview is retired.