Qwen3.6-35B-A3B runs on 16GB Mac mini

The big story here is not raw speed, it’s feasibility: a 35B-parameter open model with only ~3B active per token can be made usable on a tiny Mac, which keeps local AI from feeling gated by workstation-class hardware.

• –The posted setup uses an aggressive 4-bit quant, which is what makes the memory footprint plausible on 16GB systems
• –The shared command appears to disable GPU offload, so the result is more of a CPU-bound local inference test than a best-case Apple Silicon benchmark
• –Even at just over 6 tok/sec, this is enough for interactive use cases like private chat, agent loops, and lightweight coding help
• –For developers, the takeaway is that “open-weight frontier-ish” models are increasingly a packaging and quantization problem, not just a server-room problem
• –The result should be read as a practical anecdote, not a universal performance claim, because context size, cache settings, and backend choices will move the number a lot