llama.cpp Release b10874 Fixes Parameter Count Issue for Granite3 MoE Models
llama.cpp release b10874 resolves a bug where the parameter count for IBM Granite3 Mixture-of-Experts (MoE) models was reported as unknown. The pull request (#28632) was contributed directly by IBM engineer Aaron Teo. This fix ensures accurate model inspection and metadata reporting when executing IBM Granite3 MoE architectures locally using llama.cpp. It reflects active maintenance support from corporate AI developers for open-source local inference tooling. The patch addresses PR #28632 without altering inference algorithms or quantization logic. Compiled binary packages for release b10874 were deployed across multiple platform architectures, including macOS, Linux, Windows, and Android.
## BACKGROUND
llama.cpp is a high-performance C/C++ LLM inference framework that enables running quantized language models locally on consumer hardware. IBM Granite3 MoE refers to IBM's open Mixture-of-Experts language models engineered for low-latency tasks by activating a subset of total parameters per token.