What changed
A Hugging Face engineering prototype streams quantized mixture-of-experts weights from SSD through a proposed Transformers API, running Qwen3.8-Flash-Next with 121GB of experts on a 32GB M5 Mac and publishing code, conversion artifacts and raw comparisons.
Why it matters
Expert streaming could move very large sparse models onto commodity unified-memory systems, but it trades memory pressure for storage traffic, host round trips, conversion time and model-specific code.
What to do
Reproduce the published setup on a spare Mac with pinned PR and artifacts, record conversion time, memory, SSD reads, prefill and decode, and keep it outside production until the API and maintenance path stabilize.
Caveat
The API is still under maintainer discussion in an unmerged pull request, quantized experts use model-specific support, and the performance comparison is a single-author hardware experiment rather than an independent benchmark.