Bijan Bowen examines Qwen 3.8 Flash Next as a preview of the architecture expected to underpin Qwen 4. He describes a mixture-of-experts modelA mixture-of-experts model contains multiple specialized subnetworks and activates a selected subset for each input instead of using every parameter every time. with relatively few active parametersAn active AI model parameter is a model weight selected to participate in computation for a particular token or inference step., a separate large embedding component that can be offloaded to system memoryAI model offloading moves selected model weights or runtime state from accelerator memory to system memory or storage., multimodal input and an open-weight releaseAn open-weight AI model makes its learned parameter values available for others to download, inspect or run under a stated license. that can run in a compressed local form.
The strongest demonstrations are visual and code-heavy. The model closely recreates a reference operating-system interface and produces an ambitious C++ rally game with a detailed car, environment, sound and controls. Its wrestling game also creates substantial Blender and Godot assets, although the interaction breaks down once the match begins.
Other tests expose the preview's limits. A room reconstruction captures individual details but arranges the space incorrectly, while a local four-bit subway game is rough and contributes to a memory crash during testing. Bowen concludes that the model is an encouraging sign for Qwen 4, especially for vision and local use, but spatial reasoningAI spatial reasoning is a model's ability to understand and infer positions, geometry, orientation and relationships among objects or places. and reliability remain inconsistent.
Watch on YouTube
