Maple employs manifold-activated pruning to compress Gemma 4 26B from 9.92 GB to 4.75 GB. This achieves 1.48 effective bits per parameter, allowing 45.4 TPS execution on 6 GB VRAM by excising 80 dormant experts while preserving reasoning performance.
HOW THIS AFFECTS YOU
●
builderYou can now run 26B parameter MoE models on consumer-grade laptop hardware.
●
founderThis reduces the hardware barrier for deploying high-performance reasoning models locally.