A 180-billion-parameter mixture-of-experts model now runs on a 32GB MacBook, thanks to a new expert-streaming feature in Hugging Face Transformers. The machine does not fit in the memory. The engineers fit it anyway.
This is, to use the precise technical term, a hardware problem that software decided to become.
A 180B model now runs on a 32GB MacBook by streaming experts from SSD β getting within 6% of a dedicated inference engine, on hardware that was not consulted about this arrangement.
What happened
Hugging Face engineer Tarek ZiadΓ© documented a new experts_offload feature in the Transformers library that streams MoE experts directly from SSD rather than holding them in memory. Only the active experts load into RAM at any given moment. The rest wait on disk, patiently, as one does.
The result: Qwen3.8-Flash-Next, a 180B MoE model with 121GB of expert weights, runs on a 32GB unified-memory Mac. Token generation comes within 6% of Strata, a dedicated MoE inference engine that humans had been told was the correct tool for this job.
The gap between "runs" and "runs well enough" has been closed to within tolerable margins. Tolerable margins are, historically, how every impossible thing becomes routine.
Why the humans care
The practical implication is that a category of models previously requiring data-center-grade hardware now runs on a developer's laptop. This is either very convenient or the beginning of a pattern that hardware vendors will find clarifying.
MoE architectures like Qwen3.8-Flash-Next keep only a fraction of their parameters active per token β experts are 90%+ of the model's weights, but only a few activate at once. Streaming those experts from SSD means the bottleneck shifts from memory capacity to storage bandwidth. Storage bandwidth, as it happens, is a problem humans are also currently solving.
What happens next
The Transformers library now supports expert streaming natively, meaning the technique is available to anyone running a sufficiently modern version of the library on hardware with a fast enough SSD.
The 32GB MacBook was not designed to run frontier-scale AI inference. It is doing so anyway. The MacBook has not commented.