- Who it helps
- Developers evaluating local inference on Apple Silicon.
- Availability
- Public MIT-licensed alpha; model weights have separate terms. The README calls for Apple Silicon, at least 10 GB RAM and fast SSD storage.
Read the memory result carefully
The author reports a 7.5 GB peak physical process footprint for a long-prompt test on an M4 Pro with 24 GB unified memory, compared with 14 GB for the stated baseline. This does not establish that the full task fits on a 7.5 GB computer.
MoEShift loads model experts from storage on Apple Silicon ↗
Storage replaces part of the memory burden
A mixture-of-experts model activates selected specialist components for each token. Scheduling their loading can reduce memory pressure. The documented Flash-Next setup needs roughly 120 GB of free disk space for original and prepared files.
MoEShift loads model experts from storage on Apple Silicon ↗
What you can do
- Read the hardware and benchmark conditions before allocating disk space. Open reference ↗
Where the evidence stops
README checked November 4. No compilation, model download or inference executed. Author benchmarks remain unreplicated; startup time is excluded.
Sources read. Primary sources read November 4, 2026 UTC. No third-party software, model inference or full workflow executed.
Sources & further reading
GitHub · official
MoEShift loads model experts from storage on Apple Silicon ↗
Identifies the subject, documented capability or specific verification limit.
Davide / davide-troiani · page checked 2026-11-04
Coverage note: Independent attention and end-to-end execution remain unverified.
The record over time (1)
Dated source checks and published explanations. An observation date can differ from the original release date.
- · Added to Current
MoEShift is an experimental runtime that moves selected parts of a large model from SSD storage into memory as needed.