The llama.cpp repository has merged support for Multi-Token Prediction (MTP) in the GLM-4.5-Air model. This addition provides local inference capabilities for the latest GLM architecture via ggml.
HOW THIS AFFECTS YOU
●
builderYou can now run GLM-4.5-Air locally using llama.cpp for low-latency inference.
●
researcherEvaluate the efficiency of MTP implementation in this specific architecture.