I dumped ChatGPT Plus, Cursor and Midjourney as soon as I got my 14″ MacBook Pro with M5 Max, convinced I could run everything locally—especially after seeing Qwen 3.6 benchmarks on Apple Silicon. With a decade of backend and Docker experience, I breezed through those retention offers and swore off cloud AI. I spent my first day installing llama.cpp for general chat, Mistral via Docker for coding help, and a local Stable Diffusion build for images. By day three I hit memory bottlenecks: Qwen stuttered on longer prompts, Mistral’s inference times ballooned, and the image pipeline chewed through VRAM, forcing me to juggle swap files. On day six I struggled to integrate extensions I relied on in Cursor. The setup overhead—Python conflicts, CUDA fallbacks, random crashes—ate up more time than I saved in subscription fees. On day nine, image quality finally broke me. My local Stable Diffusion tweaks could never match Midjourney’s compositional consistency and prompt refinement tools. I threw in the towel, resubscribed to Midjourney and regained instant four-image drafts at 1024×1024. I still run Qwen locally for quick backside-of-an-envelope chats, but for anything polished—especially visuals—I’m back on the cloud.