A new architecture for Qwen4-Max’s professional ambitions
The new Qwen generation begins with a model featuring native vision, a one-million-token context window, and tool-use capabilities.
Among Alibaba’s demonstrations is an Unreal Engine game in which a cat mage battles waves of enemies. The example accompanies the release of Qwen4-Max, designed to use professional software and carry out multistep projects. The presentation also covers Blender scene creation, video production, and industrial work. These examples are provided by the vendor and do not constitute an independent assessment of success across all these activities.
The first model in the Qwen4 family is available through the QwenCloud API, as well as in Qwen Studio, QwenWork, and Qoder. The announcement includes plans to release smaller Qwen4 models with open weights, without a specific timeline. That commitment concerns future models in the family, rather than a release of the Max version’s weights. Native vision and a one-million-token context window
The mixture-of-experts architecture contains 2.4 trillion parameters in its backbone, with 71 billion activated per token. An additional 328 billion parameters are dedicated to representations of local token sequences, known as N-gram Embeddings. Its attention architecture combines Gated DeltaNet and Qwen Sparse Attention to compress the history and select relevant portions of the context.
The technical documentation confirms support for text, image, and video inputs, with a one-million-token context window. That capacity refers to the total context, rather than the maximum length of a response. Output limits are separate: 131,072 tokens without reasoning, or 262,144 tokens for reasoning and the final answer combined. From 3D scenes to video previsualization
For creative work, the examples describe using Blender for geometry, materials, lighting, and rendering. Game development involves Unity, Unreal Engine, or Godot. According to the team, the model can run a game, examine visual and interaction feedback, then adjust the camera, animations, or effects through successive tests.
Video production uses a workflow combining Qwen4 with WonderClip. The model works on storyboards, scene assets, and camera animation in previsualization environments, before rendering and editing. Preparing these parametric scenes establishes geometry, scale, spatial relationships, and the timing of actions. Cost reductions depend on the test
On Work360, its internal evaluation of professional tasks, Alibaba reports a score of 52.69, compared with 47.74 for Qwen3.8-Max. Token consumption falls by approximately 24% at the Xhigh effort level and 35% at High. The reductions approaching 50% highlighted in the announcement come from Terminal-Bench 4.0, where they reach 48% and 49%, respectively.
Lower pricing also contributes to the reported cost reduction. In the Work360 evaluation, the average cost per task falls from $4.43 to $1.67, a decrease of approximately 62%. That figure combines token consumption and API pricing; it does not guarantee equivalent savings on every project.
The QwenCloud model page lists prices of $1.50 per million input tokens, $4.50 per million output tokens, and $0.075 per million input tokens reused through implicit caching. The documentation notes that pricing varies by region. These rates should therefore be checked for the chosen endpoint. Interim releases to track long-running projects
For extended tasks, Live Delivery provides an initial usable version followed by successive releases. Users can review what has been produced, correct an error, or change their request while execution continues. Plans are then adjusted based on the results and that feedback. A public plugin supports this workflow.