MohammadHaishemKhawaja / hashyy

git clone https://github.com/MohammadHaishemKhawaja/hashyy.git

A full, unpruned 177B MoE model at 11.5 tok/s on one 12 GB GPU. Expert streaming for llama.cpp, with the measurement harness and every dead end.