MAPLE (Manifold-Activated Pruning of Latent Experts) is a new pruning technique for sub-2-BPW MoE LLMs. Its demo claims it can run Gemma 4 26B at 45 tokens per second on a 6GB GPU.