GLM-5.3-Flash (320B) on an M5 Ultra Mac Studio with llama.cpp: 61 tok/s, 262K context
retained text
Test machine: M5 ultra. 36/80. 256GB. **TL;DR** * **\~61 tok/s** generation with GLM’s built-in MTP head (probabilistic drafting, 3 draft tokens), up from **39.7** without it. * **Prompt reading is the weak spot:** \~765 tok/s at 4K tokens and \~630 tok/s at 59K. A 29K-token prompt takes 45 s. * **The full 262K context fits** with \~15 GiB to spare, and long-context recall was correct at 58.8K tokens. * **The catches:** images and MTP can’t be on at the same time, and deep context on Metal needs a fix has yet to be merged.
document fetched 2026-10-12 00:36 · canonical https://reddit.com/r/ollama/comments/1x2y2ao/glm53flash_320b_on_an_m5_ultra_mac_studio_with