Hugging Face introduces HOPE for MoE expert pruning with code and Python API

Hugging Face says its new HOPE method, or Higher-Order Pruning of Experts, prunes Mixture-of-Experts language models by modeling pairwise expert interactions instead of scoring experts independently. The team says HOPE achieves the best average rank at 50% pruning and can improve agentic coding results by up to 6.1% over the next-best pruning method. Hugging Face also released code and a compact Python API that can calibrate, solve, and prune models into smaller HuggingFace checkpoints.