llama.cpp adds support for Granite 3 Mixture-of-Experts parameter counts
September 9, 2026
A pull request in llama.cpp fixes the unknown parameter count error for Granite 3 MoE models. This update enables local inference for the Granite 3 Mixture-of-Experts architecture within the ggml runtime.
HOW THIS AFFECTS YOU
●
builderYou can now run Granite 3 MoE models locally using llama.cpp.
●
researcherThis ensures accurate parameter reporting for MoE architecture evaluation in ggml.