MAPLE: Manifold-Activated Pruning of Latent Experts

About

MAPLE prunes the 128 experts per layer of the Gemma 4 26B mixture-of-experts model down to 48 and quantizes it to a 4.75 GB GGUF file, so it fits entirely in 6 GB of laptop GPU memory. The author reports 45.4 tokens per second on an RTX 4050 versus 6.9 for the unpruned 2-bit file, which spills into system RAM. The released checkpoint is tuned for code and technical reasoning, and the repo includes scripts to profile and prune for other domains.