WiFi off, model running: what broke (and what shipped) building YOLO26 on MLX

This is a build note from the webAI YOLO26 MLX Build Challenge in May 2026, not a product pitch. The shipped result is a single-file, on-device posture attention map that runs without WiFi. The useful aspect for other builders is what went wrong during development. The team consisted of Lexi Armstrong focusing on the problem and operational constraints, while the reporter handled the engineering.

The initial architecture was a FastAPI backend with WebSocket streaming to a browser overlay, and camera capture in a ThreadPoolExecutor. However, two hard blockers killed this approach: the AVFoundation crash on macOS 26.3 and a lazy-eval deadlock in yolo-mlx Boxes proxies. The AVFoundation issue caused a SIGTRAP crash when capturing frames from a non-main thread, and the lazy-eval issue resulted in a deadlock when calling .tolist() or indexing MLX proxies after the model returned detections.

After addressing these issues, the team rewrote the code as a single synchronous loop on the main thread, which allowed for ~16 FPS at 720p on an M4 processor. This was the final product shipped to the public repository.