原帖内容

Swapped my CC's model to GLM 5.2 on my M3 Ultra 512GB and let it run for 12h on "/goal replicate Pokemon Red in HTML, make no mistakes, verify it end-to-end." This is what I got. I think I'm probably one of the few people running GLM 5.2 locally on M3 for a long-horizon task in a real CC env. My takeaway is GLM 5.2 is frontier and has good taste, but its intelligence is bottlenecked by my M3's limited inference capacity. Total duration (API): 10h 46m 4s Total duration (wall): 13h 15m 48s Total code changes: 5182 lines added, 184 lines removed Usage: 63.1k input, 127.2k output, 5.8m cache read, 0 cache write RAM 504.95 / 512.0 GB Temperature: 72°C / 162 °F Tahoe 26.5.1: unsloth/GLM-5.2-GGUF/UD-Q4_K_XL, kv cache in fp16