Alibaba has released Qwen4 Max, its new flagship AI model featuring a 2.4 trillion parameter architecture, native image understanding and a 1 million token context window. The model activates 71 billion parameters per token and scored 52.69 in Alibaba's Work360 evaluation, compared with 47.74 for Qwen3.8 Max. Alibaba says the average cost of completing a task fell from $4.43 to $1.67, a reduction of approximately 62%.
Qwen4 Max uses a Mixture of Experts architecture with 2.4 trillion total parameters, although only 71 billion are active when processing each token. This approach reduces the computing resources required for individual operations.
The model also supports image input and can process up to 1 million tokens of context. Alibaba reports that Qwen4 Max uses up to 35% fewer tokens than its predecessor in high reasoning mode, helping reduce the cost of longer tasks.
In Alibaba's Work360 benchmark, which evaluates professional tasks across 721 specialized fields, Qwen4 Max scored 52.69 points. Qwen3.8 Max reached 47.74. The average cost per task also dropped from $4.43 to $1.67. These figures come from Alibaba's own testing rather than an independent evaluation.
Qwen4 Max can work with professional software to create 3D scenes in Blender and develop games using Unity, Unreal Engine and Godot. Alibaba has also demonstrated its ability to handle manufacturing tasks, including designing production lines and preparing machining instructions for automotive parts.
In one demonstration, the model reduced the time needed to program a machining process from approximately 475 minutes to just 24 minutes. It also supports financial and legal workflows, including document analysis and the creation of reports, spreadsheets and presentations with source references.
Alibaba is also focusing on tasks that require more than a single response. Its Live Delivery feature allows Qwen4 Max to provide an early version of its work while the wider task continues. Users can review the results and request changes before the final version is ready.
The company has demonstrated these capabilities in scientific research. One project involved developing a cooling molecule for data centers through a continuous 48 hour process. Another used the model to research Alzheimer's disease and design 24 candidate molecules for further investigation.
Qwen4 Max is available through QwenCloud, with access through its API and Token Plan. Users can also access the model through Alibaba's QwenWork, Qoder and QwenStudio services. Alibaba lists cached input pricing at $0.075 per million tokens.
The reported 62% reduction applies to the average task cost in Work360. Actual API spending will depend on token consumption, caching and the type of work being performed. Alibaba says smaller open weight models from the Qwen4 family will also be released to developers, although it has not announced their release dates.



