AWS's HOPE prunes half of a MoE model's experts and, on its own test, keeps SWE-Bench Pro intact

A mixture-of-experts model only runs a few experts per token, but every expert still has to sit in memory. That is why a 35B MoE with 3B active parameters is cheap to run per token and still awkward to fit on one card. Researchers at AWS have published HOPE, a method and command-line tool that removes experts and writes out a smaller model.

What is different

Most expert-pruning methods give each expert a score from a calibration run and cut the lowest-scoring ones, layer by layer. HOPE also measures how experts work together. It builds a matrix of pairwise interactions from the same single calibration pass and picks the set to remove by solving a quadratic program. The authors show that the diagonal of that matrix reproduces the rankings of REAP, a common first-order method, so any gain over REAP comes from the pairwise terms.

The method itself dates to a September paper (arXiv 2609.18916). What is new on 9 November is the write-up and the packaged tool.

The reported results

AWS tested Qwen3.5-122B, Qwen3.5-35B and GLM-4.5-Air, with two calibration sets, at pruning rates from 10% to 50%, on the Tulu3 suite and SWE-Bench Pro. The authors say HOPE had the best average rank overall and pulled ahead at aggressive pruning.

The headline number: on Qwen3.5-35B-A3B with half the experts removed, HOPE scored 42.3 on SWE-Bench Pro against 41.8 for the unpruned model, while REAP dropped to 36.2. The authors note that pruned models sometimes edge past the original on a target task, and that other methods show the same effect, so read 42.3 as “held”, not “improved”. These are their runs, and we have not reproduced them.

Using it

The repository has three steps:

hope calibrate --model-path Qwen/Qwen3.5-35B-A3B --prompts calibration_prompts.txt ...
hope solve ...
hope prune ... --out-path ./Qwen3.5-35B-A3B-pruned-50

The output is a standard Hugging Face checkpoint, smaller on disk and in memory, and the authors say vLLM and SGLang need no model changes to load it. There is also a Python API. The package is tested on Qwen3, Qwen3-Next, Qwen3.5 and GLM-4.5. The repository’s LICENSE file carries the Apache 2.0 text.

Calibration data matters. The coding results used code-focused calibration, and the authors say that did not badly hurt the non-coding benchmarks, but your own task mix is the one to calibrate on.

Why this matters

Expert pruning is another lever next to quantization for fitting a large MoE onto the memory you have, and HOPE produces an ordinary checkpoint your existing server can load. Check the pruned model the same way you would a quantized one: compare it with KL divergence against the original, not just one benchmark. If you route many different kinds of work through one model, remember that experts can run out of capacity, and removing half of them changes that balance.