Qwen3.6-35B-A3B guide targets 32GB Macs

The real story here is not just “local AI on a Mac,” but how much system engineering it takes to make a frontier-ish coding model usable under tight memory pressure.

• –The setup leans on a quantized GGUF checkpoint plus `mmproj` support so the model can handle both code and screenshots through llama.cpp.
• –Qwen3.6-35B-A3B is positioned as an efficient open-weight MoE model with 35B total and 3B active parameters, so the value prop is density, not brute force.
• –OpenCode matters because it makes local models feel like a real coding agent instead of a toy terminal chat.
• –The author’s results are nuanced: solid on adapter-style, test-driven work, weaker on geometry-heavy UI debugging and large integration hunts.
• –The tuning advice is practical: keep context high enough to avoid collapse, but leave enough headroom for unified memory, browser tabs, and the rest of the machine.