REAM: Compressing Mixture-of-Experts LLMs

Authors: Boris Knyazev

Blog post: https://bknyaz.github.io/blog/2026/moe/

Compressed Qwen3 models 🤗: https://huggingface.co/collections/SamsungSAILMontreal/qwen3-moe

Baseline (REAP): https://github.com/CerebrasResearch/reap

We will hopefully release the code in this repo some time soon, stay tuned!