Summary
Fahd Mirza evaluates the September 2 update to Alibaba’s Qwen 3.8 Max model using a full-stack emergency dispatch application simulating New South Wales State Emergency Services incident management. The app runs on Docker with a Python backend, Postgres database, and Redis cache. The bug in question: same-priority incidents sort by smallest-to-largest affected population rather than largest-to-smallest, meaning a flood hitting 30 people sits below a minor incident in the dispatch queue — a logic error with real-world rescue consequences. Qwen 3.8 Max 0902 identifies and fixes the ordering logic on first pass, producing a correctly sorted P1-through-P4 priority queue with no targeted hints.
Mirza positions this update as a substantial leap from the previous Qwen 3.8 Max rather than a refinement, noting that the model now competes directly with frontier closed models on agentic and multimodal benchmarks rather than trailing them. Open weights are expected on Hugging Face shortly. Particular improvements are highlighted in the model’s ability to reason over large code repositories, sustain performance across long multi-step workflows, and handle multimodal inputs — all three of which are directly tested in the video.
A second evaluation presents a WhatsApp screenshot of a tangled workplace conversation and asks the model to role-play a manager choosing among three equally problematic replies. Mirza argues the model’s response demonstrates genuine risk modeling under asymmetry — identifying the one reply that minimizes information leakage across all branches — rather than defaulting to a safe non-answer. A final multilingual test spans 80-plus languages including Amharic, Sundanese, Kazakh, and Esperanto, with embedded traps to test consistency, and Mirza notes marked improvement over the model’s predecessor.
📺 Source: Fahd Mirza · Published September 02, 2026
🏷️ Format: Hands On Build
