Streaming Mixture-of-Experts Inference on Consumer Apple Silicon: A Measured Map of What Works and What Doesn't

Does a large Mixture-of-Experts (MoE) language model have to fit in RAM to run at usable speed on consumer hardware? We built a streaming inference engine for Apple Silicon — it keeps a model's attention/norm/embedding tier resident and pulls MoE expert weights off the SSD on demand through a segmented-LRU cache — and held every design decision to a live/kill gate fixed before the experiment ran. GPT-OSS-120B (117B params, ~63 GB on disk) decodes coherently at 1.04 tok/s on an 18 GB machine, a model 3.5x larger than total RAM; Qwen3-30B-A3B reaches 7.08 tok/s with a 6 GB expert cache. We report the first multi-token expert-routing predictability measurements on modern decoder MoEs, finding overlap that does not decay across an 8-token horizon. The paper's more useful half is five pre-registered negative results: shared-base + low-rank expert decomposition does not transfer to fine-grained MoEs; speculative decoding increases per-token disk traffic for expert streaming; and three engine optimizations (batched fetch, persistent expert bank, native C++/Metal fetch primitive) each miss their speed gate because the bottleneck is a per-layer CPU-GPU round-trip intrinsic to data-dependent fetch. A lightweight expert predictor cannot dodge it — the signal is real and grows with scale (+9.66 points at 120B vs ~0 at 1-3B active) but far too weak to matter. The >=3 tok/s target is unreachable for streaming MoE inference on this hardware class. We release the engine, every measurement, and every negative result.