qMLX: Maximising my AI psychosis by minmaxing my Mac Studio
Summary
A technical blog post detailing how the author fixed three bugs in a local inference stack to run Qwen 3.5 122B on a Mac Studio Ultra, achieving near-frontier performance entirely offline. The piece covers model switching, the specific bugs that broke cache and history, and the resulting gains in cache warmth and long-context throughput, powered by the qMLX fork.