- Who it helps
- Local-model developers working with limited GPU memory.
- Availability
- Public repository checked November 8. A linked model-card request failed; downloadable artifact details remain unverified.
What it does
Calibration examples identify which experts—specialized groups of model parameters—are used most. The creator keeps 48 of 128 experts per layer and builds a GGUF model file, reporting a 4.75 GB result.
What remains unverified
The calibration favors English coding and STEM tasks. Smaller size and creator-reported speed do not establish retained quality on other languages or tasks.
What you can do
- Inspect the original documentation and evaluation limits before trying it. Open reference ↗
Where the evidence stops
Documentation checked November 8, 2026 UTC. No installation or complete workflow executed.
Sources read. Primary text read; performance and operational behavior were not tested.
Sources & further reading
GitHub · implementation
MAPLE repository ↗
Creator describes calibration-based expert pruning and a smaller GGUF build. Performance was not reproduced.
AmSach · page checked 2026-11-08
Coverage note: Independent dated cross-platform discussion and complete 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
MAPLE removes less-used parts of a mixture-of-experts model to make a smaller local model.