AWS Labs open-sources HOPE, a second-order pruning method for MoE models

AWS Labs published HOPE, an open-source pruning method for mixture-of-experts models that accounts for interactions between experts and outperforms the prior best method, REAP, on agentic coding benchmarks.

Amazon's AWS Labs published HOPE (Higher-Order Pruning of Experts) on the Hugging Face blog on November 9, 2026, introducing an open-source method for pruning mixture-of-experts language models that accounts for interactions between experts. Read the Hugging Face blog post announcing HOPE for the full details.

The concrete changes

Code is published on GitHub as awslabs/HOPE, with the method detailed in a paper on arXiv (2609.18916).
At 50% pruning, HOPE reached an average rank of 1.58, compared to 2.42 for the next-best method, REAP.
HOPE delivered gains of up to 6.1% on agentic coding tasks compared to the next-best pruning method.
On the SWE-Bench Pro benchmark, a Qwen3.5-35B model pruned by 50% with HOPE scored 42.3, versus 41.8 for the unpruned baseline and 36.2 for REAP.
HOPE was tested on Qwen3.5-122B, Qwen3.5-35B and GLM-4.5-Air (106B parameters), finishing top-1 in 39% of tested conditions and top-2 in 67%.
HOPE's advantage over other pruning methods was largest at pruning rates of 40-50%.

What it means for people building with AI

Teams that self-host large open-weight mixture-of-experts models for coding agents can use HOPE to shrink those models by as much as half while losing less quality than with existing pruning methods such as REAP. That cuts the cost of running capable agentic coding models without retraining them from scratch.

Availability

The resulting pruned checkpoints are compatible with the Hugging Face transformers library, and HOPE's code and paper are published openly alongside the blog post, so teams can apply it to their own mixture-of-experts models.

Sources

Your Experts Have Partners. There's HOPE for Keeping Them Together (Hugging Face / Amazon, November 9, 2026, official)

Written with AI assistance, strictly from the sources listed above. Before it went live, automated checks verified its dates, sources and every language edition.