The author was impressed by Qwen 3.6, calling it the first local AI model that works as general intelligence. They tested both the 35B mixture-of-experts and 27B dense variants, preferring the slower but more powerful 27B version. The model successfully handled creative writing, coding tasks like hexagonal minesweeper, and practical work from simple prompts. Running locally via llama.cpp, it achieved 30 tokens per second on their MacBook M5, performing comparably to frontier models while being privately controllable and fine-tunable.