GLM-5.3-Flash (320B) runs on an M5 Ultra Mac Studio via llama.cpp, achieving ~61 token/s generation with its MTP head (probabilistic drafting, 3 draft tokens), up from ~39.7 token/s without it. Prompt processing is slower, at ~765 token/s for 4K tokens and ~630 token/s for 59K tokens, with a 29K-token prompt taking ~45 s, while the model accommodates the full 262K‑token context with ~15 GiB memory to spare and maintains correct recall up to 58.8K tokens. Limitations include the inability to enable image input and MTP simultaneously, and challenges with deep‑context performance on Metal.