Different parameter counts describe different components

Qwen reports a 2.4-trillion-parameter backbone with 71 billion parameters activated per token, plus an additional 328 billion N-gram Embedding parameters. Its architecture combines Gated DeltaNet with Qwen Sparse Attention, four parallel gated residual branches and context-based embedding lookups. Those numbers measure distinct parts of the system. The active count is not the full model size, and neither number is a demonstrated memory requirement for local deployment. The announcement does not provide a verified consumer-hardware configuration for running Max independently.

Live Delivery makes intermediate results part of the workflow

The proposed Live Delivery approach produces a usable initial result early, then expands and revises it as users provide feedback. Qwen also says it is open-sourcing a Qwen-Live-Delivery-Plugin. Showcases include Blender scenes, work in Unity, Unreal Engine and Godot, office deliverables and specialized professional tools. These are vendor demonstrations in configured environments, rather than evidence that a bare model automatically controls those applications. The distinction between model, tool access and agent harness matters when assessing whether a showcased workflow can be reproduced.

Work360 figures need their evaluation context

Qwen reports a Work360 score of 52.69, compared with 47.74 for Qwen3.8-Max. The framework’s taxonomy spans 26 top-level professional domains and 721 specialized fields. Within the company’s Work360 evaluation, average task cost reportedly falls from $4.43 to $1.67. That is not a universal service price or an independent forecast for every customer. Task mix, reasoning effort, available tools and token accounting all influence the result. A cost reduction in a benchmark should be distinguished from the invoice for a particular production workflow.

Cloud access is separate from the open-model roadmap

The official page names QwenCloud through a Token Plan or API, alongside QwenWork, Qoder and QwenStudio. Its promise of smaller open-source models does not confer an open-weight license on Max, so local inference and fine-tuning rights should not be assumed. Likewise, a million-token context window is an input-capacity specification, not a guarantee that every detail will be retrieved correctly. Teams considering integration have a concrete new flagship to evaluate, but should check actual product access, the tool environment and performance on their own long-running tasks before treating vendor demonstrations as transferable results.