TL;DR: llama.cpp release b10603 adds multi-token prediction (MTP) support for the GLM-4.5-Air model, enabling faster local inference.
Summary: The b10603 release of llama.cpp now supports MTP (multi-token prediction) in GLM-4.5-Air via PR #26534. MTP lets the model predict multiple tokens per forward pass, reducing generation latency. The release also ships updated iOS XCFramework and UI builds.
Why it matters: AI builders running GLM-4.5-Air locally can update llama.cpp to get faster token generation, especially on constrained hardware. Try enabling MTP and benchmark generation speed against the previous release.
Source: github_releases