Hugging Face introduces HOPE for efficient MoE model pruning

hugging face has introduced hope, a second-order method for pruning mixture-of-experts moE models. hope considers pairwise expert interactions, unlike existing first-order methods that score experts independently. in experiments, hope achieves the best average rank at 50% pruning, with gains of up to 6.1% on agentic coding tasks over the next-best method. this development helps address the memory bottleneck in large language models, enabling more efficient deployment. hope can be used with quantization and avoids latency issues associated with dynamic loading and off-loading of experts.