Amazon researchers have introduced HOPE, short for Higher-Order Pruning of Experts, a method for shrinking Mixture-of-Experts (MoE) language models by permanently removing the least useful expert networks. MoE models route each token to a small subset of experts, but every expert must still occupy memory at inference time. The article cites Qwen3.5-122B, which has 256 experts in each of its 48 MoE layers. Expert networks account for about 95% of its roughly 250 GB memory footprint, yet only eight experts per layer are active for any given token. Existing pruning methods score each expert independently and delete the lowest scorers. HOPE instead accounts for pairwise interactions, based on the authors' observation that experts form stable groups that are frequently activated together. Removing an expert that a frequent partner relies on can degrade the model more than its individual score suggests. The authors say HOPE achieves the best average rank of any tested method at 50% pruning, with gains of up to 6.1% on agentic coding tasks over the next-best pruning method. The technique requires no retraining and can be combined with quantization, according to the post. These figures come from the authors' own paper, and the excerpt reviewed does not include the full benchmark breakdown or the implementation walkthrough. For Australian businesses and professionals, this changes nothing immediately for teams using hosted AI APIs, since the provider carries the serving costs. For organisations that self-host large open-weight MoE models, pruning could in principle reduce GPU memory requirements and hardware spend, but this is research-stage work worth evaluating in a test environment rather than deploying. Any pruned model that processes personal information remains subject to the Privacy Act 1988 obligations that already apply to that data.