AmSach/maple: Manifold-Activated Pruning of Latent Experts for Sub-2-BPW MoE LLMs

Independent researcher Aman Sachan has released MAPLE, a domain-agnostic pruning pipeline for Mixture-of-Experts language models that targets sub-2-BPW structural compression. By profiling live token activations and excising dormant multilingual experts, MAPLE successfully compresses the Gemma 4 26B-A4B model from 9.92 GB to 4.75 GB. This optimization enables 100% in-VRAM execution on 6 GB consumer GPUs with a significant speedup in token throughput. Open article →