Summary
Fahd Mirza tests Qwen 3.8 Max — Alibaba’s most capable model to date and the first Max-class release to have its weights open-sourced — putting it through a live multi-bug debugging challenge using the Hermes agent framework. The model is a 22.4 trillion parameter mixture-of-experts architecture designed for long-horizon autonomous work, with Qwen citing demonstrations of it running unattended for over 10 days to build software, reproduce and beat research papers, and climb live competition leaderboards.
The hands-on test involves a broken FastAPI + SQLite + Redis full-stack NYC subway board application with five chained bugs planted across the codebase — including one silent bug with no error output. Qwen 3.8 Max, operating through Hermes agent with no hints, identifies and fixes all five, including the silent pub/sub stream bug. A parallel test on Qwen’s hosted chat platform generates a complex 15-scene 3D WebGL page from a single prompt, further demonstrating the model’s ability to hold large, detailed specifications coherently.
Notably, Mirza flags an announcement embedded in the Qwen release notes: an upcoming open-weights release of Qwen 3.8 27B under an Apache 2.0 license on Hugging Face — a follow-up to the 3.6 27B model he considers among the best dense open-source models available. For developers and researchers tracking the open-weights frontier, that forthcoming release may be the more consequential news from this video.
📺 Source: Fahd Mirza · Published August 03, 2026
🏷️ Format: Review
