Qwen3.5-35B-A3B Pushes Mac Memory Limits

The short answer is that 24GB may be fine for general development, but it is not the comfortable tier for local 35B-class LLM hobby work. If you want room for larger context, fewer compromises, and less babysitting, 48GB is the safer buy.

• –Qwen3.5-35B-A3B is a 35B-total MoE model with 3B active parameters, so it is efficient for its class but still not "small" in memory terms.
• –The thread's comments are consistent: 24GB is described as barely enough, while 48GB avoids the yellow-zone memory pressure people hit on Apple Silicon.
• –For local inference, unified memory matters as much as raw CPU/GPU speed because KV cache and context quickly eat the available headroom.
• –If the goal is occasional experimentation, 24GB can work with aggressive quantization and shorter contexts; if the goal is a serious local LLM setup, 48GB is the practical minimum.