A local GLM-5.2 serving writeup reports soloheaven hitting about 18 tokens per second versus 3 to 8 on stock mlx_lm.server, with session KV cache reuse on follow-up calls. The useful caveat is that this is one operator's M3 Ultra measurement, plus a strict=False loader patch and a 0.0.0.0 proxy that deserve careful review before anyone copies the setup blindly.