MAPLE: Manifold-Activated Pruning of Latent Experts

We present MAPLE (Manifold-Activated Pruning of Latent Experts), a post-training structural compression method for fine-grained Mixture-of-Experts (MoE) large language models that achieves sub-2-bit-per-weight effective parameter density while retaining high task-level semantic fidelity. In fine-grained MoE models trained on expansive multilingual corpora, a substantial fraction of routed expert capacity remains dormant for any specific deployment subdomain. We exploit this structural phenomenon by profiling the empirical expert activation manifold during native forward passes over a representative domain corpus. Each expert is scored via a composite metric combining activation frequency and input feature energy: se,l = Ce,l · √(σ̄e,l + ε). Applied to Gemma 4 26B-A4B (128 routed experts, 30 layers, 9.92 GB in IQ2_XXS quantization), MAPLE excises 80 dormant experts per layer to assemble an optimized 48-expert container of 4.75 GB (effective 1.48 bits per parameter, 2.09× compression). The pruned model achieves 100% GPU VRAM residency on a consumer 6 GB laptop GPU (RTX 4050), accelerating generation by 6.57× from 6.9 to 45.4 tokens/second by eliminating PCIe memory-paging bottlenecks. Standardized evaluation across code synthesis, data structure implementation, scientific deduction, and formal logic demonstrates a 93.8% mean keyword fidelity retention. We document three subtle failure modes overcome during design: proxy-embedding routing estimation, cascaded normalization suppression, and DRY sampling interference.