Bijan Bowen evaluates the open-weight Qwen 3.8 Max mixture-of-experts model across coding, web design, browser use, game creation, 3D work and physical reasoning. The model has 2.4 trillion total parameters with 95 billion active per token, giving it a large capacity while limiting inference cost relative to a dense model of the same size.
The strongest results were a detailed C++ skateboarding game set in New York and a polished wedding website. The model followed long design instructions, generated substantial working code and often produced usable interfaces without extensive repair.
Performance was less reliable on tasks involving physical constraints, CAD-style geometry, browser interaction and some 3D scenes. The full test also took a long time and cost about 32 dollars, so the review concludes that the model is capable but still uneven across modalities and not automatically the best choice for every workload.
Watch the original on YouTube