llama.cpp adds GLM-5.3-Flash MoE model support

llama.cpp is gaining support for GLM-5.3-Flash (also called GLM-5-Next), Zhipu's large mixture-of-experts model at roughly 321 billion parameters, aimed at people who want to run it locally through the usual GGUF stack.

Early testers say the integration still rebuilds its decode graph on every token, which costs on the order of 7% of decode throughput. Real speed on multi-GPU boxes is dominated less by that overhead than by quantization and how CPU-offloaded experts are scheduled. Q8_0 keeps most of the model in host RAM and pays full dequant on the CPU each step; denser Unsloth-style quants cut resident size sharply. On CPU-heavy MoE runs, thread counts in a narrow band around the mid-20s to low-30s outperform the common half-of-nproc default. One dual RTX 3090 plus EPYC setup reported much higher token rates after those changes, though some of the peak figures depended on extra expert caching not in the base support work.

Multimodal use is not ready. Loading a vision projector fails with an unknown glm5v projector type, so image input does not work until that path is taught the new type.