Summary
Fahd Mirza tests Qwen 3.8 Max preview within minutes of its announcement, putting Alibaba’s latest flagship model through demanding prompts before the inference servers become overloaded. The model is described as a 2.4 trillion parameter release — currently available in preview on the Qwen platform with an open-weights release planned — which Alibaba positions as competitive with leading frontier models and second only to GPT-5, a claim Mirza sets out to verify rather than accept.
The opening test asks Qwen 3.8 to generate a self-contained HTML canvas simulation combining over twelve global railway systems — including the Trans-Siberian, Shinkansen, and Metro — with regional biomes, block signaling, day/night cycles, rain effects, animated birds, and working speed controls. The result runs in-browser and impresses Mirza enough to call it better than GPT-5 on this task. A second test challenges the model to write original poetry in nearly 80 languages, each in the distinctive voice of a recognized poet from that language, without quoting or paraphrasing any real published works.
Mirza also tests multimodal reasoning by showing the model an image of a traffic island received as a divorce settlement, asking what can be built there. The model correctly identifies it as a traffic island and responds with a blend of legal clarity and deadpan humor. For developers tracking the Qwen family — which Mirza notes has consistently contributed more to the open-source community than DeepSeek — this video offers an unfiltered first look at what may be Alibaba’s most capable model to date.
📺 Source: Fahd Mirza · Published July 19, 2026
🏷️ Format: Review
